6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "asm/register.hpp"
26 #include "ci/ciObjArray.hpp"
27 #include "ci/ciUtilities.hpp"
28 #include "classfile/javaClasses.hpp"
29 #include "compiler/compileLog.hpp"
30 #include "gc/shared/barrierSet.hpp"
31 #include "gc/shared/c2/barrierSetC2.hpp"
32 #include "interpreter/interpreter.hpp"
33 #include "memory/resourceArea.hpp"
34 #include "opto/addnode.hpp"
35 #include "opto/castnode.hpp"
36 #include "opto/convertnode.hpp"
37 #include "opto/graphKit.hpp"
38 #include "opto/idealKit.hpp"
39 #include "opto/intrinsicnode.hpp"
40 #include "opto/locknode.hpp"
41 #include "opto/machnode.hpp"
42 #include "opto/opaquenode.hpp"
43 #include "opto/parse.hpp"
44 #include "opto/rootnode.hpp"
45 #include "opto/runtime.hpp"
46 #include "opto/subtypenode.hpp"
47 #include "runtime/deoptimization.hpp"
48 #include "runtime/sharedRuntime.hpp"
49 #include "utilities/bitMap.inline.hpp"
50 #include "utilities/growableArray.hpp"
51 #include "utilities/powerOfTwo.hpp"
52
53 //----------------------------GraphKit-----------------------------------------
54 // Main utility constructor.
55 GraphKit::GraphKit(JVMState* jvms)
56 : Phase(Phase::Parser),
57 _env(C->env()),
58 _gvn(*C->initial_gvn()),
59 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
60 {
61 _exceptions = jvms->map()->next_exception();
62 if (_exceptions != nullptr) jvms->map()->set_next_exception(nullptr);
63 set_jvms(jvms);
64 }
65
66 // Private constructor for parser.
67 GraphKit::GraphKit()
68 : Phase(Phase::Parser),
69 _env(C->env()),
70 _gvn(*C->initial_gvn()),
71 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
72 {
73 _exceptions = nullptr;
74 set_map(nullptr);
75 DEBUG_ONLY(_sp = -99);
76 DEBUG_ONLY(set_bci(-99));
77 }
78
79
80
81 //---------------------------clean_stack---------------------------------------
82 // Clear away rubbish from the stack area of the JVM state.
83 // This destroys any arguments that may be waiting on the stack.
328 }
329 static inline void add_one_req(Node* dstphi, Node* src) {
330 assert(is_hidden_merge(dstphi), "must be a special merge node");
331 assert(!is_hidden_merge(src), "must not be a special merge node");
332 dstphi->add_req(src);
333 }
334
335 //-----------------------combine_exception_states------------------------------
336 // This helper function combines exception states by building phis on a
337 // specially marked state-merging region. These regions and phis are
338 // untransformed, and can build up gradually. The region is marked by
339 // having a control input of its exception map, rather than null. Such
340 // regions do not appear except in this function, and in use_exception_state.
341 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
342 if (failing_internal()) {
343 return; // dying anyway...
344 }
345 JVMState* ex_jvms = ex_map->_jvms;
346 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
347 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
348 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
349 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
350 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
351 assert(ex_map->req() == phi_map->req(), "matching maps");
352 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
353 Node* hidden_merge_mark = root();
354 Node* region = phi_map->control();
355 MergeMemNode* phi_mem = phi_map->merged_memory();
356 MergeMemNode* ex_mem = ex_map->merged_memory();
357 if (region->in(0) != hidden_merge_mark) {
358 // The control input is not (yet) a specially-marked region in phi_map.
359 // Make it so, and build some phis.
360 region = new RegionNode(2);
361 _gvn.set_type(region, Type::CONTROL);
362 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
363 region->init_req(1, phi_map->control());
364 phi_map->set_control(region);
365 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
366 record_for_igvn(io_phi);
367 _gvn.set_type(io_phi, Type::ABIO);
368 phi_map->set_i_o(io_phi);
856 if (PrintMiscellaneous && (Verbose || WizardMode)) {
857 tty->print_cr("Zombie local %d: ", local);
858 jvms->dump();
859 }
860 return false;
861 }
862 }
863 }
864 return true;
865 }
866
867 #endif //ASSERT
868
869 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens.
870 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
871 ciMethod* cur_method = jvms->method();
872 int cur_bci = jvms->bci();
873 if (cur_method != nullptr && cur_bci != InvocationEntryBci) {
874 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
875 return Interpreter::bytecode_should_reexecute(code) ||
876 (is_anewarray && code == Bytecodes::_multianewarray);
877 // Reexecute _multianewarray bytecode which was replaced with
878 // sequence of [a]newarray. See Parse::do_multianewarray().
879 //
880 // Note: interpreter should not have it set since this optimization
881 // is limited by dimensions and guarded by flag so in some cases
882 // multianewarray() runtime calls will be generated and
883 // the bytecode should not be reexecutes (stack will not be reset).
884 } else {
885 return false;
886 }
887 }
888
889 // Helper function for adding JVMState and debug information to node
890 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
891 // Add the safepoint edges to the call (or other safepoint).
892
893 // Make sure dead locals are set to top. This
894 // should help register allocation time and cut down on the size
895 // of the deoptimization information.
896 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
947 }
948
949 // Presize the call:
950 DEBUG_ONLY(uint non_debug_edges = call->req());
951 call->add_req_batch(top(), youngest_jvms->debug_depth());
952 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
953
954 // Set up edges so that the call looks like this:
955 // Call [state:] ctl io mem fptr retadr
956 // [parms:] parm0 ... parmN
957 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
958 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
959 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
960 // Note that caller debug info precedes callee debug info.
961
962 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
963 uint debug_ptr = call->req();
964
965 // Loop over the map input edges associated with jvms, add them
966 // to the call node, & reset all offsets to match call node array.
967 for (JVMState* in_jvms = youngest_jvms; in_jvms != nullptr; ) {
968 uint debug_end = debug_ptr;
969 uint debug_start = debug_ptr - in_jvms->debug_size();
970 debug_ptr = debug_start; // back up the ptr
971
972 uint p = debug_start; // walks forward in [debug_start, debug_end)
973 uint j, k, l;
974 SafePointNode* in_map = in_jvms->map();
975 out_jvms->set_map(call);
976
977 if (can_prune_locals) {
978 assert(in_jvms->method() == out_jvms->method(), "sanity");
979 // If the current throw can reach an exception handler in this JVMS,
980 // then we must keep everything live that can reach that handler.
981 // As a quick and dirty approximation, we look for any handlers at all.
982 if (in_jvms->method()->has_exception_handlers()) {
983 can_prune_locals = false;
984 }
985 }
986
987 // Add the Locals
988 k = in_jvms->locoff();
989 l = in_jvms->loc_size();
990 out_jvms->set_locoff(p);
991 if (!can_prune_locals) {
992 for (j = 0; j < l; j++)
993 call->set_req(p++, in_map->in(k+j));
994 } else {
995 p += l; // already set to top above by add_req_batch
996 }
997
998 // Add the Expression Stack
999 k = in_jvms->stkoff();
1000 l = in_jvms->sp();
1001 out_jvms->set_stkoff(p);
1002 if (!can_prune_locals) {
1003 for (j = 0; j < l; j++)
1004 call->set_req(p++, in_map->in(k+j));
1005 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
1006 // Divide stack into {S0,...,S1}, where S0 is set to top.
1007 uint s1 = stack_slots_not_pruned;
1008 stack_slots_not_pruned = 0; // for next iteration
1009 if (s1 > l) s1 = l;
1010 uint s0 = l - s1;
1011 p += s0; // skip the tops preinstalled by add_req_batch
1012 for (j = s0; j < l; j++)
1013 call->set_req(p++, in_map->in(k+j));
1014 } else {
1015 p += l; // already set to top above by add_req_batch
1016 }
1017
1018 // Add the Monitors
1019 k = in_jvms->monoff();
1020 l = in_jvms->mon_size();
1021 out_jvms->set_monoff(p);
1022 for (j = 0; j < l; j++)
1023 call->set_req(p++, in_map->in(k+j));
1024
1025 // Copy any scalar object fields.
1026 k = in_jvms->scloff();
1027 l = in_jvms->scl_size();
1028 out_jvms->set_scloff(p);
1029 for (j = 0; j < l; j++)
1030 call->set_req(p++, in_map->in(k+j));
1031
1032 // Finish the new jvms.
1033 out_jvms->set_endoff(p);
1034
1035 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
1036 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
1037 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
1038 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
1039 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
1040 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
1041
1042 // Update the two tail pointers in parallel.
1043 out_jvms = out_jvms->caller();
1044 in_jvms = in_jvms->caller();
1045 }
1046
1047 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
1048
1049 // Test the correctness of JVMState::debug_xxx accessors:
1050 assert(call->jvms()->debug_start() == non_debug_edges, "");
1051 assert(call->jvms()->debug_end() == call->req(), "");
1052 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1053 }
1054
1055 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1056 Bytecodes::Code code = java_bc();
1057 if (code == Bytecodes::_wide) {
1058 code = method()->java_code_at_bci(bci() + 1);
1059 }
1060
1061 if (code != Bytecodes::_illegal) {
1062 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1198 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1199 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1200 return _gvn.transform( new AndLNode(conv, mask) );
1201 }
1202
1203 Node* GraphKit::ConvL2I(Node* offset) {
1204 // short-circuit a common case
1205 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1206 if (offset_con != (jlong)Type::OffsetBot) {
1207 return intcon((int) offset_con);
1208 }
1209 return _gvn.transform( new ConvL2INode(offset));
1210 }
1211
1212 //-------------------------load_object_klass-----------------------------------
1213 Node* GraphKit::load_object_klass(Node* obj) {
1214 // Special-case a fresh allocation to avoid building nodes:
1215 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1216 if (akls != nullptr) return akls;
1217 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1218 return _gvn.transform(LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1219 }
1220
1221 //-------------------------load_array_length-----------------------------------
1222 Node* GraphKit::load_array_length(Node* array) {
1223 // Special-case a fresh allocation to avoid building nodes:
1224 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array);
1225 Node *alen;
1226 if (alloc == nullptr) {
1227 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1228 alen = _gvn.transform( new LoadRangeNode(nullptr, immutable_memory(), r_adr, TypeInt::POS));
1229 } else {
1230 alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false);
1231 }
1232 return alen;
1233 }
1234
1235 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc,
1236 const TypeOopPtr* oop_type,
1237 bool replace_length_in_map) {
1238 Node* length = alloc->Ideal_length();
1247 replace_in_map(length, ccast);
1248 }
1249 return ccast;
1250 }
1251 }
1252 return length;
1253 }
1254
1255 //------------------------------do_null_check----------------------------------
1256 // Helper function to do a null pointer check. Returned value is
1257 // the incoming address with null casted away. You are allowed to use the
1258 // not-null value only if you are control dependent on the test.
1259 #ifndef PRODUCT
1260 extern uint explicit_null_checks_inserted,
1261 explicit_null_checks_elided;
1262 #endif
1263 Node* GraphKit::null_check_common(Node* value, BasicType type,
1264 // optional arguments for variations:
1265 bool assert_null,
1266 Node* *null_control,
1267 bool speculative) {
1268 assert(!assert_null || null_control == nullptr, "not both at once");
1269 if (stopped()) return top();
1270 NOT_PRODUCT(explicit_null_checks_inserted++);
1271
1272 // Construct null check
1273 Node *chk = nullptr;
1274 switch(type) {
1275 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1276 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1277 case T_ARRAY : // fall through
1278 type = T_OBJECT; // simplify further tests
1279 case T_OBJECT : {
1280 const Type *t = _gvn.type( value );
1281
1282 const TypeOopPtr* tp = t->isa_oopptr();
1283 if (tp != nullptr && !tp->is_loaded()
1284 // Only for do_null_check, not any of its siblings:
1285 && !assert_null && null_control == nullptr) {
1286 // Usually, any field access or invocation on an unloaded oop type
1287 // will simply fail to link, since the statically linked class is
1288 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1289 // the static class is loaded but the sharper oop type is not.
1290 // Rather than checking for this obscure case in lots of places,
1291 // we simply observe that a null check on an unloaded class
1355 }
1356 Node *oldcontrol = control();
1357 set_control(cfg);
1358 Node *res = cast_not_null(value);
1359 set_control(oldcontrol);
1360 NOT_PRODUCT(explicit_null_checks_elided++);
1361 return res;
1362 }
1363 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1364 if (cfg == nullptr) break; // Quit at region nodes
1365 depth++;
1366 }
1367 }
1368
1369 //-----------
1370 // Branch to failure if null
1371 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1372 Deoptimization::DeoptReason reason;
1373 if (assert_null) {
1374 reason = Deoptimization::reason_null_assert(speculative);
1375 } else if (type == T_OBJECT) {
1376 reason = Deoptimization::reason_null_check(speculative);
1377 } else {
1378 reason = Deoptimization::Reason_div0_check;
1379 }
1380 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1381 // ciMethodData::has_trap_at will return a conservative -1 if any
1382 // must-be-null assertion has failed. This could cause performance
1383 // problems for a method after its first do_null_assert failure.
1384 // Consider using 'Reason_class_check' instead?
1385
1386 // To cause an implicit null check, we set the not-null probability
1387 // to the maximum (PROB_MAX). For an explicit check the probability
1388 // is set to a smaller value.
1389 if (null_control != nullptr || too_many_traps(reason)) {
1390 // probability is less likely
1391 ok_prob = PROB_LIKELY_MAG(3);
1392 } else if (!assert_null &&
1393 (ImplicitNullCheckThreshold > 0) &&
1394 method() != nullptr &&
1395 (method()->method_data()->trap_count(reason)
1429 }
1430
1431 if (assert_null) {
1432 // Cast obj to null on this path.
1433 replace_in_map(value, zerocon(type));
1434 return zerocon(type);
1435 }
1436
1437 // Cast obj to not-null on this path, if there is no null_control.
1438 // (If there is a null_control, a non-null value may come back to haunt us.)
1439 if (type == T_OBJECT) {
1440 Node* cast = cast_not_null(value, false);
1441 if (null_control == nullptr || (*null_control) == top())
1442 replace_in_map(value, cast);
1443 value = cast;
1444 }
1445
1446 return value;
1447 }
1448
1449
1450 //------------------------------cast_not_null----------------------------------
1451 // Cast obj to not-null on this path
1452 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1453 const Type *t = _gvn.type(obj);
1454 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1455 // Object is already not-null?
1456 if( t == t_not_null ) return obj;
1457
1458 Node* cast = new CastPPNode(control(), obj,t_not_null);
1459 cast = _gvn.transform( cast );
1460
1461 // Scan for instances of 'obj' in the current JVM mapping.
1462 // These instances are known to be not-null after the test.
1463 if (do_replace_in_map)
1464 replace_in_map(obj, cast);
1465
1466 return cast; // Return casted value
1467 }
1468
1469 // Sometimes in intrinsics, we implicitly know an object is not null
1470 // (there's no actual null check) so we can cast it to not null. In
1471 // the course of optimizations, the input to the cast can become null.
1472 // In that case that data path will die and we need the control path
1473 // to become dead as well to keep the graph consistent. So we have to
1474 // add a check for null for which one branch can't be taken. It uses
1475 // an OpaqueNotNull node that will cause the check to be removed after loop
1476 // opts so the test goes away and the compiled code doesn't execute a
1477 // useless check.
1478 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1479 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) {
1480 return value;
1481 }
1482 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1483 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1484 Node* opaq = _gvn.transform(new OpaqueNotNullNode(C, tst));
1485 IfNode* iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1486 _gvn.set_type(iff, iff->Value(&_gvn));
1487 if (!tst->is_Con()) {
1488 record_for_igvn(iff);
1561 // These are layered on top of the factory methods in LoadNode and StoreNode,
1562 // and integrate with the parser's memory state and _gvn engine.
1563 //
1564
1565 // factory methods in "int adr_idx"
1566 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1567 MemNode::MemOrd mo,
1568 LoadNode::ControlDependency control_dependency,
1569 bool require_atomic_access,
1570 bool unaligned,
1571 bool mismatched,
1572 bool unsafe,
1573 uint8_t barrier_data) {
1574 int adr_idx = C->get_alias_index(_gvn.type(adr)->isa_ptr());
1575 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1576 const TypePtr* adr_type = nullptr; // debug-mode-only argument
1577 DEBUG_ONLY(adr_type = C->get_adr_type(adr_idx));
1578 Node* mem = memory(adr_idx);
1579 Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data);
1580 ld = _gvn.transform(ld);
1581 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1582 // Improve graph before escape analysis and boxing elimination.
1583 record_for_igvn(ld);
1584 if (ld->is_DecodeN()) {
1585 // Also record the actual load (LoadN) in case ld is DecodeN. In some
1586 // rare corner cases, ld->in(1) can be something other than LoadN (e.g.,
1587 // a Phi). Recording such cases is still perfectly sound, but may be
1588 // unnecessary and result in some minor IGVN overhead.
1589 record_for_igvn(ld->in(1));
1590 }
1591 }
1592 return ld;
1593 }
1594
1595 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1596 MemNode::MemOrd mo,
1597 bool require_atomic_access,
1598 bool unaligned,
1599 bool mismatched,
1600 bool unsafe,
1614 if (unsafe) {
1615 st->as_Store()->set_unsafe_access();
1616 }
1617 st->as_Store()->set_barrier_data(barrier_data);
1618 st = _gvn.transform(st);
1619 set_memory(st, adr_idx);
1620 // Back-to-back stores can only remove intermediate store with DU info
1621 // so push on worklist for optimizer.
1622 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1623 record_for_igvn(st);
1624
1625 return st;
1626 }
1627
1628 Node* GraphKit::access_store_at(Node* obj,
1629 Node* adr,
1630 const TypePtr* adr_type,
1631 Node* val,
1632 const Type* val_type,
1633 BasicType bt,
1634 DecoratorSet decorators) {
1635 // Transformation of a value which could be null pointer (CastPP #null)
1636 // could be delayed during Parse (for example, in adjust_map_after_if()).
1637 // Execute transformation here to avoid barrier generation in such case.
1638 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1639 val = _gvn.makecon(TypePtr::NULL_PTR);
1640 }
1641
1642 if (stopped()) {
1643 return top(); // Dead path ?
1644 }
1645
1646 assert(val != nullptr, "not dead path");
1647
1648 C2AccessValuePtr addr(adr, adr_type);
1649 C2AccessValue value(val, val_type);
1650 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr);
1651 if (access.is_raw()) {
1652 return _barrier_set->BarrierSetC2::store_at(access, value);
1653 } else {
1654 return _barrier_set->store_at(access, value);
1655 }
1656 }
1657
1658 Node* GraphKit::access_load_at(Node* obj, // containing obj
1659 Node* adr, // actual address to store val at
1660 const TypePtr* adr_type,
1661 const Type* val_type,
1662 BasicType bt,
1663 DecoratorSet decorators) {
1664 if (stopped()) {
1665 return top(); // Dead path ?
1666 }
1667
1668 C2AccessValuePtr addr(adr, adr_type);
1669 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr);
1670 if (access.is_raw()) {
1671 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1672 } else {
1673 return _barrier_set->load_at(access, val_type);
1674 }
1675 }
1676
1677 Node* GraphKit::access_load(Node* adr, // actual address to load val at
1678 const Type* val_type,
1679 BasicType bt,
1680 DecoratorSet decorators) {
1681 if (stopped()) {
1682 return top(); // Dead path ?
1683 }
1684
1685 C2AccessValuePtr addr(adr, adr->bottom_type()->is_ptr());
1686 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, nullptr, addr);
1687 if (access.is_raw()) {
1688 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1689 } else {
1754 Node* new_val,
1755 const Type* value_type,
1756 BasicType bt,
1757 DecoratorSet decorators) {
1758 C2AccessValuePtr addr(adr, adr_type);
1759 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1760 if (access.is_raw()) {
1761 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1762 } else {
1763 return _barrier_set->atomic_add_at(access, new_val, value_type);
1764 }
1765 }
1766
1767 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1768 return _barrier_set->clone(this, src, dst, size, is_array);
1769 }
1770
1771 //-------------------------array_element_address-------------------------
1772 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1773 const TypeInt* sizetype, Node* ctrl) {
1774 uint shift = exact_log2(type2aelembytes(elembt));
1775 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1776
1777 // short-circuit a common case (saves lots of confusing waste motion)
1778 jint idx_con = find_int_con(idx, -1);
1779 if (idx_con >= 0) {
1780 intptr_t offset = header + ((intptr_t)idx_con << shift);
1781 return basic_plus_adr(ary, offset);
1782 }
1783
1784 // must be correct type for alignment purposes
1785 Node* base = basic_plus_adr(ary, header);
1786 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1787 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1788 return basic_plus_adr(ary, base, scale);
1789 }
1790
1791 //-------------------------load_array_element-------------------------
1792 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) {
1793 const Type* elemtype = arytype->elem();
1794 BasicType elembt = elemtype->array_element_basic_type();
1795 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1796 if (elembt == T_NARROWOOP) {
1797 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1798 }
1799 Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt,
1800 IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0));
1801 return ld;
1802 }
1803
1804 //-------------------------set_arguments_for_java_call-------------------------
1805 // Arguments (pre-popped from the stack) are taken from the JVMS.
1806 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1807 // Add the call arguments:
1808 uint nargs = call->method()->arg_size();
1809 for (uint i = 0; i < nargs; i++) {
1810 Node* arg = argument(i);
1811 call->init_req(i + TypeFunc::Parms, arg);
1812 }
1813 }
1814
1815 //---------------------------set_edges_for_java_call---------------------------
1816 // Connect a newly created call into the current JVMS.
1817 // A return value node (if any) is returned from set_edges_for_java_call.
1818 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1819
1820 // Add the predefined inputs:
1821 call->init_req( TypeFunc::Control, control() );
1822 call->init_req( TypeFunc::I_O , i_o() );
1823 call->init_req( TypeFunc::Memory , reset_memory() );
1824 call->init_req( TypeFunc::FramePtr, frameptr() );
1825 call->init_req( TypeFunc::ReturnAdr, top() );
1826
1827 add_safepoint_edges(call, must_throw);
1828
1829 Node* xcall = _gvn.transform(call);
1830
1831 if (xcall == top()) {
1832 set_control(top());
1833 return;
1834 }
1835 assert(xcall == call, "call identity is stable");
1836
1837 // Re-use the current map to produce the result.
1838
1839 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1840 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1841 set_all_memory_call(xcall, separate_io_proj);
1842
1843 //return xcall; // no need, caller already has it
1844 }
1845
1846 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
1847 if (stopped()) return top(); // maybe the call folded up?
1848
1849 // Capture the return value, if any.
1850 Node* ret;
1851 if (call->method() == nullptr ||
1852 call->method()->return_type()->basic_type() == T_VOID)
1853 ret = top();
1854 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1855
1856 // Note: Since any out-of-line call can produce an exception,
1857 // we always insert an I_O projection from the call into the result.
1858
1859 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
1860
1861 if (separate_io_proj) {
1862 // The caller requested separate projections be used by the fall
1863 // through and exceptional paths, so replace the projections for
1864 // the fall through path.
1865 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1866 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1867 }
1868 return ret;
1869 }
1870
1871 //--------------------set_predefined_input_for_runtime_call--------------------
1872 // Reading and setting the memory state is way conservative here.
1873 // The real problem is that I am not doing real Type analysis on memory,
1874 // so I cannot distinguish card mark stores from other stores. Across a GC
1875 // point the Store Barrier and the card mark memory has to agree. I cannot
1876 // have a card mark store and its barrier split across the GC point from
1877 // either above or below. Here I get that to happen by reading ALL of memory.
1878 // A better answer would be to separate out card marks from other memory.
1879 // For now, return the input memory state, so that it can be reused
1880 // after the call, if this call has restricted memory effects.
1881 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
1882 // Set fixed predefined input arguments
1883 Node* memory = reset_memory();
1884 Node* m = narrow_mem == nullptr ? memory : narrow_mem;
1885 call->init_req( TypeFunc::Control, control() );
1886 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1887 call->init_req( TypeFunc::Memory, m ); // may gc ptrs
1938 if (use->is_MergeMem()) {
1939 wl.push(use);
1940 }
1941 }
1942 }
1943
1944 // Replace the call with the current state of the kit.
1945 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes, bool do_asserts) {
1946 JVMState* ejvms = nullptr;
1947 if (has_exceptions()) {
1948 ejvms = transfer_exceptions_into_jvms();
1949 }
1950
1951 ReplacedNodes replaced_nodes = map()->replaced_nodes();
1952 ReplacedNodes replaced_nodes_exception;
1953 Node* ex_ctl = top();
1954
1955 SafePointNode* final_state = stop();
1956
1957 // Find all the needed outputs of this call
1958 CallProjections callprojs;
1959 call->extract_projections(&callprojs, true, do_asserts);
1960
1961 Unique_Node_List wl;
1962 Node* init_mem = call->in(TypeFunc::Memory);
1963 Node* final_mem = final_state->in(TypeFunc::Memory);
1964 Node* final_ctl = final_state->in(TypeFunc::Control);
1965 Node* final_io = final_state->in(TypeFunc::I_O);
1966
1967 // Replace all the old call edges with the edges from the inlining result
1968 if (callprojs.fallthrough_catchproj != nullptr) {
1969 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1970 }
1971 if (callprojs.fallthrough_memproj != nullptr) {
1972 if (final_mem->is_MergeMem()) {
1973 // Parser's exits MergeMem was not transformed but may be optimized
1974 final_mem = _gvn.transform(final_mem);
1975 }
1976 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
1977 add_mergemem_users_to_worklist(wl, final_mem);
1978 }
1979 if (callprojs.fallthrough_ioproj != nullptr) {
1980 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
1981 }
1982
1983 // Replace the result with the new result if it exists and is used
1984 if (callprojs.resproj != nullptr && result != nullptr) {
1985 C->gvn_replace_by(callprojs.resproj, result);
1986 }
1987
1988 if (ejvms == nullptr) {
1989 // No exception edges to simply kill off those paths
1990 if (callprojs.catchall_catchproj != nullptr) {
1991 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1992 }
1993 if (callprojs.catchall_memproj != nullptr) {
1994 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1995 }
1996 if (callprojs.catchall_ioproj != nullptr) {
1997 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1998 }
1999 // Replace the old exception object with top
2000 if (callprojs.exobj != nullptr) {
2001 C->gvn_replace_by(callprojs.exobj, C->top());
2002 }
2003 } else {
2004 GraphKit ekit(ejvms);
2005
2006 // Load my combined exception state into the kit, with all phis transformed:
2007 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
2008 replaced_nodes_exception = ex_map->replaced_nodes();
2009
2010 Node* ex_oop = ekit.use_exception_state(ex_map);
2011
2012 if (callprojs.catchall_catchproj != nullptr) {
2013 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
2014 ex_ctl = ekit.control();
2015 }
2016 if (callprojs.catchall_memproj != nullptr) {
2017 Node* ex_mem = ekit.reset_memory();
2018 C->gvn_replace_by(callprojs.catchall_memproj, ex_mem);
2019 add_mergemem_users_to_worklist(wl, ex_mem);
2020 }
2021 if (callprojs.catchall_ioproj != nullptr) {
2022 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
2023 }
2024
2025 // Replace the old exception object with the newly created one
2026 if (callprojs.exobj != nullptr) {
2027 C->gvn_replace_by(callprojs.exobj, ex_oop);
2028 }
2029 }
2030
2031 // Disconnect the call from the graph
2032 call->disconnect_inputs(C);
2033 C->gvn_replace_by(call, C->top());
2034
2035 // Clean up any MergeMems that feed other MergeMems since the
2036 // optimizer doesn't like that.
2037 while (wl.size() > 0) {
2038 _gvn.transform(wl.pop());
2039 }
2040
2041 if (callprojs.fallthrough_catchproj != nullptr && !final_ctl->is_top() && do_replaced_nodes) {
2042 replaced_nodes.apply(C, final_ctl);
2043 }
2044 if (!ex_ctl->is_top() && do_replaced_nodes) {
2045 replaced_nodes_exception.apply(C, ex_ctl);
2046 }
2047 }
2048
2049
2050 //------------------------------increment_counter------------------------------
2051 // for statistics: increment a VM counter by 1
2052
2053 void GraphKit::increment_counter(address counter_addr) {
2054 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2055 increment_counter(adr1);
2056 }
2057
2058 void GraphKit::increment_counter(Node* counter_addr) {
2059 Node* ctrl = control();
2060 Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, MemNode::unordered);
2061 Node* incr = _gvn.transform(new AddLNode(cnt, _gvn.longcon(1)));
2221 *
2222 * @param n node that the type applies to
2223 * @param exact_kls type from profiling
2224 * @param maybe_null did profiling see null?
2225 *
2226 * @return node with improved type
2227 */
2228 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2229 const Type* current_type = _gvn.type(n);
2230 assert(UseTypeSpeculation, "type speculation must be on");
2231
2232 const TypePtr* speculative = current_type->speculative();
2233
2234 // Should the klass from the profile be recorded in the speculative type?
2235 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2236 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls, Type::trust_interfaces);
2237 const TypeOopPtr* xtype = tklass->as_instance_type();
2238 assert(xtype->klass_is_exact(), "Should be exact");
2239 // Any reason to believe n is not null (from this profiling or a previous one)?
2240 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2241 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2242 // record the new speculative type's depth
2243 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2244 speculative = speculative->with_inline_depth(jvms()->depth());
2245 } else if (current_type->would_improve_ptr(ptr_kind)) {
2246 // Profiling report that null was never seen so we can change the
2247 // speculative type to non null ptr.
2248 if (ptr_kind == ProfileAlwaysNull) {
2249 speculative = TypePtr::NULL_PTR;
2250 } else {
2251 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2252 const TypePtr* ptr = TypePtr::NOTNULL;
2253 if (speculative != nullptr) {
2254 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2255 } else {
2256 speculative = ptr;
2257 }
2258 }
2259 }
2260
2261 if (speculative != current_type->speculative()) {
2262 // Build a type with a speculative type (what we think we know
2263 // about the type but will need a guard when we use it)
2264 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2265 // We're changing the type, we need a new CheckCast node to carry
2266 // the new type. The new type depends on the control: what
2267 // profiling tells us is only valid from here as far as we can
2268 // tell.
2269 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2270 cast = _gvn.transform(cast);
2271 replace_in_map(n, cast);
2272 n = cast;
2273 }
2274
2275 return n;
2276 }
2277
2278 /**
2279 * Record profiling data from receiver profiling at an invoke with the
2280 * type system so that it can propagate it (speculation)
2281 *
2282 * @param n receiver node
2283 *
2284 * @return node with improved type
2285 */
2286 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2287 if (!UseTypeSpeculation) {
2288 return n;
2289 }
2290 ciKlass* exact_kls = profile_has_unique_klass();
2291 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2292 if ((java_bc() == Bytecodes::_checkcast ||
2293 java_bc() == Bytecodes::_instanceof ||
2294 java_bc() == Bytecodes::_aastore) &&
2295 method()->method_data()->is_mature()) {
2296 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2297 if (data != nullptr) {
2298 if (!data->as_BitData()->null_seen()) {
2299 ptr_kind = ProfileNeverNull;
2300 } else {
2301 assert(data->is_ReceiverTypeData(), "bad profile data type");
2302 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2303 uint i = 0;
2304 for (; i < call->row_limit(); i++) {
2305 ciKlass* receiver = call->receiver(i);
2306 if (receiver != nullptr) {
2307 break;
2308 }
2309 }
2310 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2311 }
2312 }
2313 }
2314 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2315 }
2316
2317 /**
2318 * Record profiling data from argument profiling at an invoke with the
2319 * type system so that it can propagate it (speculation)
2320 *
2321 * @param dest_method target method for the call
2322 * @param bc what invoke bytecode is this?
2323 */
2324 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2325 if (!UseTypeSpeculation) {
2326 return;
2327 }
2328 const TypeFunc* tf = TypeFunc::make(dest_method);
2329 int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2330 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2331 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2332 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2333 if (is_reference_type(targ->basic_type())) {
2334 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2335 ciKlass* better_type = nullptr;
2336 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2337 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2338 }
2339 i++;
2340 }
2341 }
2342 }
2343
2344 /**
2345 * Record profiling data from parameter profiling at an invoke with
2346 * the type system so that it can propagate it (speculation)
2347 */
2348 void GraphKit::record_profiled_parameters_for_speculation() {
2349 if (!UseTypeSpeculation) {
2350 return;
2351 }
2352 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2472 // The first null ends the list.
2473 Node* parm0, Node* parm1,
2474 Node* parm2, Node* parm3,
2475 Node* parm4, Node* parm5,
2476 Node* parm6, Node* parm7) {
2477 assert(call_addr != nullptr, "must not call null targets");
2478
2479 // Slow-path call
2480 bool is_leaf = !(flags & RC_NO_LEAF);
2481 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2482 if (call_name == nullptr) {
2483 assert(!is_leaf, "must supply name for leaf");
2484 call_name = OptoRuntime::stub_name(call_addr);
2485 }
2486 CallNode* call;
2487 if (!is_leaf) {
2488 call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type);
2489 } else if (flags & RC_NO_FP) {
2490 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2491 } else if (flags & RC_VECTOR){
2492 uint num_bits = call_type->range()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte;
2493 call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits);
2494 } else {
2495 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2496 }
2497
2498 // The following is similar to set_edges_for_java_call,
2499 // except that the memory effects of the call are restricted to AliasIdxRaw.
2500
2501 // Slow path call has no side-effects, uses few values
2502 bool wide_in = !(flags & RC_NARROW_MEM);
2503 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2504
2505 Node* prev_mem = nullptr;
2506 if (wide_in) {
2507 prev_mem = set_predefined_input_for_runtime_call(call);
2508 } else {
2509 assert(!wide_out, "narrow in => narrow out");
2510 Node* narrow_mem = memory(adr_type);
2511 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2512 }
2513
2514 // Hook each parm in order. Stop looking at the first null.
2515 if (parm0 != nullptr) { call->init_req(TypeFunc::Parms+0, parm0);
2516 if (parm1 != nullptr) { call->init_req(TypeFunc::Parms+1, parm1);
2517 if (parm2 != nullptr) { call->init_req(TypeFunc::Parms+2, parm2);
2518 if (parm3 != nullptr) { call->init_req(TypeFunc::Parms+3, parm3);
2519 if (parm4 != nullptr) { call->init_req(TypeFunc::Parms+4, parm4);
2520 if (parm5 != nullptr) { call->init_req(TypeFunc::Parms+5, parm5);
2521 if (parm6 != nullptr) { call->init_req(TypeFunc::Parms+6, parm6);
2522 if (parm7 != nullptr) { call->init_req(TypeFunc::Parms+7, parm7);
2523 /* close each nested if ===> */ } } } } } } } }
2524 assert(call->in(call->req()-1) != nullptr, "must initialize all parms");
2525
2526 if (!is_leaf) {
2527 // Non-leaves can block and take safepoints:
2528 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2529 }
2530 // Non-leaves can throw exceptions:
2531 if (has_io) {
2532 call->set_req(TypeFunc::I_O, i_o());
2533 }
2534
2535 if (flags & RC_UNCOMMON) {
2536 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2537 // (An "if" probability corresponds roughly to an unconditional count.
2538 // Sort of.)
2539 call->set_cnt(PROB_UNLIKELY_MAG(4));
2540 }
2541
2542 Node* c = _gvn.transform(call);
2543 assert(c == call, "cannot disappear");
2544
2552
2553 if (has_io) {
2554 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2555 }
2556 return call;
2557
2558 }
2559
2560 // i2b
2561 Node* GraphKit::sign_extend_byte(Node* in) {
2562 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24)));
2563 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24)));
2564 }
2565
2566 // i2s
2567 Node* GraphKit::sign_extend_short(Node* in) {
2568 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16)));
2569 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16)));
2570 }
2571
2572 //------------------------------merge_memory-----------------------------------
2573 // Merge memory from one path into the current memory state.
2574 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2575 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2576 Node* old_slice = mms.force_memory();
2577 Node* new_slice = mms.memory2();
2578 if (old_slice != new_slice) {
2579 PhiNode* phi;
2580 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2581 if (mms.is_empty()) {
2582 // clone base memory Phi's inputs for this memory slice
2583 assert(old_slice == mms.base_memory(), "sanity");
2584 phi = PhiNode::make(region, nullptr, Type::MEMORY, mms.adr_type(C));
2585 _gvn.set_type(phi, Type::MEMORY);
2586 for (uint i = 1; i < phi->req(); i++) {
2587 phi->init_req(i, old_slice->in(i));
2588 }
2589 } else {
2590 phi = old_slice->as_Phi(); // Phi was generated already
2591 }
2648 gvn.transform(iff);
2649 if (!bol->is_Con()) gvn.record_for_igvn(iff);
2650 return iff;
2651 }
2652
2653 //-------------------------------gen_subtype_check-----------------------------
2654 // Generate a subtyping check. Takes as input the subtype and supertype.
2655 // Returns 2 values: sets the default control() to the true path and returns
2656 // the false path. Only reads invariant memory; sets no (visible) memory.
2657 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2658 // but that's not exposed to the optimizer. This call also doesn't take in an
2659 // Object; if you wish to check an Object you need to load the Object's class
2660 // prior to coming here.
2661 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn,
2662 ciMethod* method, int bci) {
2663 Compile* C = gvn.C;
2664 if ((*ctrl)->is_top()) {
2665 return C->top();
2666 }
2667
2668 // Fast check for identical types, perhaps identical constants.
2669 // The types can even be identical non-constants, in cases
2670 // involving Array.newInstance, Object.clone, etc.
2671 if (subklass == superklass)
2672 return C->top(); // false path is dead; no test needed.
2673
2674 if (gvn.type(superklass)->singleton()) {
2675 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
2676 const TypeKlassPtr* subk = gvn.type(subklass)->is_klassptr();
2677
2678 // In the common case of an exact superklass, try to fold up the
2679 // test before generating code. You may ask, why not just generate
2680 // the code and then let it fold up? The answer is that the generated
2681 // code will necessarily include null checks, which do not always
2682 // completely fold away. If they are also needless, then they turn
2683 // into a performance loss. Example:
2684 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2685 // Here, the type of 'fa' is often exact, so the store check
2686 // of fa[1]=x will fold up, without testing the nullness of x.
2687 //
2688 // At macro expansion, we would have already folded the SubTypeCheckNode
2689 // being expanded here because we always perform the static sub type
2690 // check in SubTypeCheckNode::sub() regardless of whether
2691 // StressReflectiveCode is set or not. We can therefore skip this
2692 // static check when StressReflectiveCode is on.
2693 switch (C->static_subtype_check(superk, subk)) {
2694 case Compile::SSC_always_false:
2695 {
2696 Node* always_fail = *ctrl;
2697 *ctrl = gvn.C->top();
2698 return always_fail;
2699 }
2700 case Compile::SSC_always_true:
2701 return C->top();
2702 case Compile::SSC_easy_test:
2703 {
2704 // Just do a direct pointer compare and be done.
2705 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2706 *ctrl = gvn.transform(new IfTrueNode(iff));
2707 return gvn.transform(new IfFalseNode(iff));
2708 }
2709 case Compile::SSC_full_test:
2710 break;
2711 default:
2712 ShouldNotReachHere();
2713 }
2714 }
2715
2716 // %%% Possible further optimization: Even if the superklass is not exact,
2717 // if the subklass is the unique subtype of the superklass, the check
2718 // will always succeed. We could leave a dependency behind to ensure this.
2719
2720 // First load the super-klass's check-offset
2721 Node *p1 = gvn.transform(new AddPNode(superklass, superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2722 Node* m = C->immutable_memory();
2723 Node *chk_off = gvn.transform(new LoadINode(nullptr, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2724 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2725 const TypeInt* chk_off_t = chk_off->Value(&gvn)->isa_int();
2726 int chk_off_con = (chk_off_t != nullptr && chk_off_t->is_con()) ? chk_off_t->get_con() : cacheoff_con;
2727 bool might_be_cache = (chk_off_con == cacheoff_con);
2728
2729 // Load from the sub-klass's super-class display list, or a 1-word cache of
2730 // the secondary superclass list, or a failing value with a sentinel offset
2731 // if the super-klass is an interface or exceptionally deep in the Java
2732 // hierarchy and we have to scan the secondary superclass list the hard way.
2733 // Worst-case type is a little odd: null is allowed as a result (usually
2734 // klass loads can never produce a null).
2735 Node *chk_off_X = chk_off;
2736 #ifdef _LP64
2737 chk_off_X = gvn.transform(new ConvI2LNode(chk_off_X));
2738 #endif
2739 Node *p2 = gvn.transform(new AddPNode(subklass,subklass,chk_off_X));
2740 // For some types like interfaces the following loadKlass is from a 1-word
2741 // cache which is mutable so can't use immutable memory. Other
2742 // types load from the super-class display table which is immutable.
2743 Node *kmem = C->immutable_memory();
2744 // secondary_super_cache is not immutable but can be treated as such because:
2745 // - no ideal node writes to it in a way that could cause an
2746 // incorrect/missed optimization of the following Load.
2747 // - it's a cache so, worse case, not reading the latest value
2758
2759 // Gather the various success & failures here
2760 RegionNode* r_not_subtype = new RegionNode(3);
2761 gvn.record_for_igvn(r_not_subtype);
2762 RegionNode* r_ok_subtype = new RegionNode(4);
2763 gvn.record_for_igvn(r_ok_subtype);
2764
2765 // If we might perform an expensive check, first try to take advantage of profile data that was attached to the
2766 // SubTypeCheck node
2767 if (might_be_cache && method != nullptr && VM_Version::profile_all_receivers_at_type_check()) {
2768 ciCallProfile profile = method->call_profile_at_bci(bci);
2769 float total_prob = 0;
2770 for (int i = 0; profile.has_receiver(i); ++i) {
2771 float prob = profile.receiver_prob(i);
2772 total_prob += prob;
2773 }
2774 if (total_prob * 100. >= TypeProfileSubTypeCheckCommonThreshold) {
2775 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
2776 for (int i = 0; profile.has_receiver(i); ++i) {
2777 ciKlass* klass = profile.receiver(i);
2778 const TypeKlassPtr* klass_t = TypeKlassPtr::make(klass);
2779 Compile::SubTypeCheckResult result = C->static_subtype_check(superk, klass_t);
2780 if (result != Compile::SSC_always_true && result != Compile::SSC_always_false) {
2781 continue;
2782 }
2783 float prob = profile.receiver_prob(i);
2784 ConNode* klass_node = gvn.makecon(klass_t);
2785 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, klass_node, BoolTest::eq, prob, gvn, T_ADDRESS);
2786 Node* iftrue = gvn.transform(new IfTrueNode(iff));
2787
2788 if (result == Compile::SSC_always_true) {
2789 r_ok_subtype->add_req(iftrue);
2790 } else {
2791 assert(result == Compile::SSC_always_false, "");
2792 r_not_subtype->add_req(iftrue);
2793 }
2794 *ctrl = gvn.transform(new IfFalseNode(iff));
2795 }
2796 }
2797 }
2808 // subklass. In this case we need exactly the 1 test above and we can
2809 // return those results immediately.
2810 if (!might_be_cache) {
2811 Node* not_subtype_ctrl = *ctrl;
2812 *ctrl = iftrue1; // We need exactly the 1 test above
2813 PhaseIterGVN* igvn = gvn.is_IterGVN();
2814 if (igvn != nullptr) {
2815 igvn->remove_globally_dead_node(r_ok_subtype);
2816 igvn->remove_globally_dead_node(r_not_subtype);
2817 }
2818 return not_subtype_ctrl;
2819 }
2820
2821 r_ok_subtype->init_req(1, iftrue1);
2822
2823 // Check for immediate negative hit. Happens roughly 11% of the time (which
2824 // is roughly 63% of the remaining cases). Test to see if the loaded
2825 // check-offset points into the subklass display list or the 1-element
2826 // cache. If it points to the display (and NOT the cache) and the display
2827 // missed then it's not a subtype.
2828 Node *cacheoff = gvn.intcon(cacheoff_con);
2829 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2830 r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2)));
2831 *ctrl = gvn.transform(new IfFalseNode(iff2));
2832
2833 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2834 // No performance impact (too rare) but allows sharing of secondary arrays
2835 // which has some footprint reduction.
2836 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2837 r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3)));
2838 *ctrl = gvn.transform(new IfFalseNode(iff3));
2839
2840 // -- Roads not taken here: --
2841 // We could also have chosen to perform the self-check at the beginning
2842 // of this code sequence, as the assembler does. This would not pay off
2843 // the same way, since the optimizer, unlike the assembler, can perform
2844 // static type analysis to fold away many successful self-checks.
2845 // Non-foldable self checks work better here in second position, because
2846 // the initial primary superclass check subsumes a self-check for most
2847 // types. An exception would be a secondary type like array-of-interface,
2848 // which does not appear in its own primary supertype display.
2849 // Finally, we could have chosen to move the self-check into the
2850 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2851 // dependent manner. But it is worthwhile to have the check here,
2852 // where it can be perhaps be optimized. The cost in code space is
2853 // small (register compare, branch).
2854
2855 // Now do a linear scan of the secondary super-klass array. Again, no real
2856 // performance impact (too rare) but it's gotta be done.
2857 // Since the code is rarely used, there is no penalty for moving it
2858 // out of line, and it can only improve I-cache density.
2859 // The decision to inline or out-of-line this final check is platform
2860 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2861 Node* psc = gvn.transform(
2862 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2863
2864 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2865 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
2866 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
2867
2868 // Return false path; set default control to true path.
2869 *ctrl = gvn.transform(r_ok_subtype);
2870 return gvn.transform(r_not_subtype);
2871 }
2872
2873 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
2874 bool expand_subtype_check = C->post_loop_opts_phase(); // macro node expansion is over
2875 if (expand_subtype_check) {
2876 MergeMemNode* mem = merged_memory();
2877 Node* ctrl = control();
2878 Node* subklass = obj_or_subklass;
2879 if (!_gvn.type(obj_or_subklass)->isa_klassptr()) {
2880 subklass = load_object_klass(obj_or_subklass);
2881 }
2882
2883 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn, method(), bci());
2884 set_control(ctrl);
2885 return n;
2886 }
2887
2888 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass, method(), bci()));
2889 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
2890 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2891 set_control(_gvn.transform(new IfTrueNode(iff)));
2892 return _gvn.transform(new IfFalseNode(iff));
2893 }
2894
2895 // Profile-driven exact type check:
2896 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2897 float prob,
2898 Node* *casted_receiver) {
2899 assert(!klass->is_interface(), "no exact type check on interfaces");
2900
2901 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces);
2902 Node* recv_klass = load_object_klass(receiver);
2903 Node* want_klass = makecon(tklass);
2904 Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass));
2905 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
2906 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2907 set_control( _gvn.transform(new IfTrueNode (iff)));
2908 Node* fail = _gvn.transform(new IfFalseNode(iff));
2909
2910 if (!stopped()) {
2911 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
2912 const TypeOopPtr* recvx_type = tklass->as_instance_type();
2913 assert(recvx_type->klass_is_exact(), "");
2914
2915 if (!receiver_type->higher_equal(recvx_type)) { // ignore redundant casts
2916 // Subsume downstream occurrences of receiver with a cast to
2917 // recv_xtype, since now we know what the type will be.
2918 Node* cast = new CheckCastPPNode(control(), receiver, recvx_type);
2919 (*casted_receiver) = _gvn.transform(cast);
2920 assert(!(*casted_receiver)->is_top(), "that path should be unreachable");
2921 // (User must make the replace_in_map call.)
2922 }
2923 }
2924
2925 return fail;
2926 }
2927
2928 //------------------------------subtype_check_receiver-------------------------
2929 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
2930 Node** casted_receiver) {
2931 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces)->try_improve();
2932 Node* want_klass = makecon(tklass);
2933
2934 Node* slow_ctl = gen_subtype_check(receiver, want_klass);
2935
2936 // Ignore interface type information until interface types are properly tracked.
2937 if (!stopped() && !klass->is_interface()) {
2938 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
2939 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
2940 if (!receiver_type->higher_equal(recv_type)) { // ignore redundant casts
2941 Node* cast = new CheckCastPPNode(control(), receiver, recv_type);
2942 (*casted_receiver) = _gvn.transform(cast);
2943 }
2944 }
2945
2946 return slow_ctl;
2947 }
2948
2949 //------------------------------seems_never_null-------------------------------
2950 // Use null_seen information if it is available from the profile.
2951 // If we see an unexpected null at a type check we record it and force a
2952 // recompile; the offending check will be recompiled to handle nulls.
2953 // If we see several offending BCIs, then all checks in the
2954 // method will be recompiled.
2955 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
2956 speculating = !_gvn.type(obj)->speculative_maybe_null();
2957 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
2958 if (UncommonNullCast // Cutout for this technique
2959 && obj != null() // And not the -Xcomp stupid case?
2960 && !too_many_traps(reason)
2961 ) {
2962 if (speculating) {
3031
3032 //------------------------maybe_cast_profiled_receiver-------------------------
3033 // If the profile has seen exactly one type, narrow to exactly that type.
3034 // Subsequent type checks will always fold up.
3035 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3036 const TypeKlassPtr* require_klass,
3037 ciKlass* spec_klass,
3038 bool safe_for_replace) {
3039 if (!UseTypeProfile || !TypeProfileCasts) return nullptr;
3040
3041 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != nullptr);
3042
3043 // Make sure we haven't already deoptimized from this tactic.
3044 if (too_many_traps_or_recompiles(reason))
3045 return nullptr;
3046
3047 // (No, this isn't a call, but it's enough like a virtual call
3048 // to use the same ciMethod accessor to get the profile info...)
3049 // If we have a speculative type use it instead of profiling (which
3050 // may not help us)
3051 ciKlass* exact_kls = spec_klass == nullptr ? profile_has_unique_klass() : spec_klass;
3052 if (exact_kls != nullptr) {// no cast failures here
3053 if (require_klass == nullptr ||
3054 C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls, Type::trust_interfaces)) == Compile::SSC_always_true) {
3055 // If we narrow the type to match what the type profile sees or
3056 // the speculative type, we can then remove the rest of the
3057 // cast.
3058 // This is a win, even if the exact_kls is very specific,
3059 // because downstream operations, such as method calls,
3060 // will often benefit from the sharper type.
3061 Node* exact_obj = not_null_obj; // will get updated in place...
3062 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3063 &exact_obj);
3064 { PreserveJVMState pjvms(this);
3065 set_control(slow_ctl);
3066 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3067 }
3068 if (safe_for_replace) {
3069 replace_in_map(not_null_obj, exact_obj);
3070 }
3071 return exact_obj;
3161 // If not_null_obj is dead, only null-path is taken
3162 if (stopped()) { // Doing instance-of on a null?
3163 set_control(null_ctl);
3164 return intcon(0);
3165 }
3166 region->init_req(_null_path, null_ctl);
3167 phi ->init_req(_null_path, intcon(0)); // Set null path value
3168 if (null_ctl == top()) {
3169 // Do this eagerly, so that pattern matches like is_diamond_phi
3170 // will work even during parsing.
3171 assert(_null_path == PATH_LIMIT-1, "delete last");
3172 region->del_req(_null_path);
3173 phi ->del_req(_null_path);
3174 }
3175
3176 // Do we know the type check always succeed?
3177 bool known_statically = false;
3178 if (_gvn.type(superklass)->singleton()) {
3179 const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr();
3180 const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type();
3181 if (subk->is_loaded()) {
3182 int static_res = C->static_subtype_check(superk, subk);
3183 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3184 }
3185 }
3186
3187 if (!known_statically) {
3188 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3189 // We may not have profiling here or it may not help us. If we
3190 // have a speculative type use it to perform an exact cast.
3191 ciKlass* spec_obj_type = obj_type->speculative_type();
3192 if (spec_obj_type != nullptr || (ProfileDynamicTypes && data != nullptr)) {
3193 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, nullptr, spec_obj_type, safe_for_replace);
3194 if (stopped()) { // Profile disagrees with this path.
3195 set_control(null_ctl); // Null is the only remaining possibility.
3196 return intcon(0);
3197 }
3198 if (cast_obj != nullptr) {
3199 not_null_obj = cast_obj;
3200 }
3201 }
3217 record_for_igvn(region);
3218
3219 // If we know the type check always succeeds then we don't use the
3220 // profiling data at this bytecode. Don't lose it, feed it to the
3221 // type system as a speculative type.
3222 if (safe_for_replace) {
3223 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3224 replace_in_map(obj, casted_obj);
3225 }
3226
3227 return _gvn.transform(phi);
3228 }
3229
3230 //-------------------------------gen_checkcast---------------------------------
3231 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3232 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3233 // uncommon-trap paths work. Adjust stack after this call.
3234 // If failure_control is supplied and not null, it is filled in with
3235 // the control edge for the cast failure. Otherwise, an appropriate
3236 // uncommon trap or exception is thrown.
3237 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
3238 Node* *failure_control) {
3239 kill_dead_locals(); // Benefit all the uncommon traps
3240 const TypeKlassPtr* klass_ptr_type = _gvn.type(superklass)->is_klassptr();
3241 const TypeKlassPtr* improved_klass_ptr_type = klass_ptr_type->try_improve();
3242 const TypeOopPtr* toop = improved_klass_ptr_type->cast_to_exactness(false)->as_instance_type();
3243
3244 // Fast cutout: Check the case that the cast is vacuously true.
3245 // This detects the common cases where the test will short-circuit
3246 // away completely. We do this before we perform the null check,
3247 // because if the test is going to turn into zero code, we don't
3248 // want a residual null check left around. (Causes a slowdown,
3249 // for example, in some objArray manipulations, such as a[i]=a[j].)
3250 if (improved_klass_ptr_type->singleton()) {
3251 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3252 if (objtp != nullptr) {
3253 switch (C->static_subtype_check(improved_klass_ptr_type, objtp->as_klass_type())) {
3254 case Compile::SSC_always_true:
3255 // If we know the type check always succeed then we don't use
3256 // the profiling data at this bytecode. Don't lose it, feed it
3257 // to the type system as a speculative type.
3258 return record_profiled_receiver_for_speculation(obj);
3259 case Compile::SSC_always_false:
3260 // It needs a null check because a null will *pass* the cast check.
3261 // A non-null value will always produce an exception.
3262 if (!objtp->maybe_null()) {
3263 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3264 Deoptimization::DeoptReason reason = is_aastore ?
3265 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3266 builtin_throw(reason);
3267 return top();
3268 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3269 return null_assert(obj);
3270 }
3271 break; // Fall through to full check
3272 default:
3273 break;
3274 }
3275 }
3276 }
3277
3278 ciProfileData* data = nullptr;
3279 bool safe_for_replace = false;
3280 if (failure_control == nullptr) { // use MDO in regular case only
3281 assert(java_bc() == Bytecodes::_aastore ||
3282 java_bc() == Bytecodes::_checkcast,
3283 "interpreter profiles type checks only for these BCs");
3284 data = method()->method_data()->bci_to_data(bci());
3285 safe_for_replace = true;
3286 }
3287
3288 // Make the merge point
3289 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3290 RegionNode* region = new RegionNode(PATH_LIMIT);
3291 Node* phi = new PhiNode(region, toop);
3292 C->set_has_split_ifs(true); // Has chance for split-if optimization
3293
3294 // Use null-cast information if it is available
3295 bool speculative_not_null = false;
3296 bool never_see_null = ((failure_control == nullptr) // regular case only
3297 && seems_never_null(obj, data, speculative_not_null));
3298
3299 // Null check; get casted pointer; set region slot 3
3300 Node* null_ctl = top();
3301 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3302
3303 // If not_null_obj is dead, only null-path is taken
3304 if (stopped()) { // Doing instance-of on a null?
3305 set_control(null_ctl);
3306 return null();
3307 }
3308 region->init_req(_null_path, null_ctl);
3309 phi ->init_req(_null_path, null()); // Set null path value
3310 if (null_ctl == top()) {
3311 // Do this eagerly, so that pattern matches like is_diamond_phi
3312 // will work even during parsing.
3313 assert(_null_path == PATH_LIMIT-1, "delete last");
3314 region->del_req(_null_path);
3315 phi ->del_req(_null_path);
3316 }
3317
3318 Node* cast_obj = nullptr;
3319 if (improved_klass_ptr_type->klass_is_exact()) {
3320 // The following optimization tries to statically cast the speculative type of the object
3321 // (for example obtained during profiling) to the type of the superklass and then do a
3322 // dynamic check that the type of the object is what we expect. To work correctly
3323 // for checkcast and aastore the type of superklass should be exact.
3324 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3325 // We may not have profiling here or it may not help us. If we have
3326 // a speculative type use it to perform an exact cast.
3327 ciKlass* spec_obj_type = obj_type->speculative_type();
3328 if (spec_obj_type != nullptr || data != nullptr) {
3329 cast_obj = maybe_cast_profiled_receiver(not_null_obj, improved_klass_ptr_type, spec_obj_type, safe_for_replace);
3330 if (cast_obj != nullptr) {
3331 if (failure_control != nullptr) // failure is now impossible
3332 (*failure_control) = top();
3333 // adjust the type of the phi to the exact klass:
3334 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3335 }
3336 }
3337 }
3338
3339 if (cast_obj == nullptr) {
3340 // Generate the subtype check
3341 Node* improved_superklass = superklass;
3342 if (improved_klass_ptr_type != klass_ptr_type && improved_klass_ptr_type->singleton()) {
3343 improved_superklass = makecon(improved_klass_ptr_type);
3344 }
3345 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, improved_superklass);
3346
3347 // Plug in success path into the merge
3348 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3349 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3350 if (failure_control == nullptr) {
3351 if (not_subtype_ctrl != top()) { // If failure is possible
3352 PreserveJVMState pjvms(this);
3353 set_control(not_subtype_ctrl);
3354 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3355 Deoptimization::DeoptReason reason = is_aastore ?
3356 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3357 builtin_throw(reason);
3358 }
3359 } else {
3360 (*failure_control) = not_subtype_ctrl;
3361 }
3362 }
3363
3364 region->init_req(_obj_path, control());
3365 phi ->init_req(_obj_path, cast_obj);
3366
3367 // A merge of null or Casted-NotNull obj
3368 Node* res = _gvn.transform(phi);
3369
3370 // Note I do NOT always 'replace_in_map(obj,result)' here.
3371 // if( tk->klass()->can_be_primary_super() )
3372 // This means that if I successfully store an Object into an array-of-String
3373 // I 'forget' that the Object is really now known to be a String. I have to
3374 // do this because we don't have true union types for interfaces - if I store
3375 // a Baz into an array-of-Interface and then tell the optimizer it's an
3376 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3377 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3378 // replace_in_map( obj, res );
3379
3380 // Return final merged results
3381 set_control( _gvn.transform(region) );
3382 record_for_igvn(region);
3383
3384 return record_profiled_receiver_for_speculation(res);
3385 }
3386
3387 //------------------------------next_monitor-----------------------------------
3388 // What number should be given to the next monitor?
3389 int GraphKit::next_monitor() {
3390 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3391 int next = current + C->sync_stack_slots();
3392 // Keep the toplevel high water mark current:
3393 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3394 return current;
3395 }
3396
3397 //------------------------------insert_mem_bar---------------------------------
3398 // Memory barrier to avoid floating things around
3399 // The membar serves as a pinch point between both control and all memory slices.
3400 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3401 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3402 mb->init_req(TypeFunc::Control, control());
3403 mb->init_req(TypeFunc::Memory, reset_memory());
3404 Node* membar = _gvn.transform(mb);
3432 }
3433 Node* membar = _gvn.transform(mb);
3434 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3435 if (alias_idx == Compile::AliasIdxBot) {
3436 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3437 } else {
3438 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3439 }
3440 return membar;
3441 }
3442
3443 //------------------------------shared_lock------------------------------------
3444 // Emit locking code.
3445 FastLockNode* GraphKit::shared_lock(Node* obj) {
3446 // bci is either a monitorenter bc or InvocationEntryBci
3447 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3448 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3449
3450 if( !GenerateSynchronizationCode )
3451 return nullptr; // Not locking things?
3452 if (stopped()) // Dead monitor?
3453 return nullptr;
3454
3455 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3456
3457 // Box the stack location
3458 Node* box = new BoxLockNode(next_monitor());
3459 // Check for bailout after new BoxLockNode
3460 if (failing()) { return nullptr; }
3461 box = _gvn.transform(box);
3462 Node* mem = reset_memory();
3463
3464 FastLockNode * flock = _gvn.transform(new FastLockNode(nullptr, obj, box) )->as_FastLock();
3465
3466 // Add monitor to debug info for the slow path. If we block inside the
3467 // slow path and de-opt, we need the monitor hanging around
3468 map()->push_monitor( flock );
3469
3470 const TypeFunc *tf = LockNode::lock_type();
3471 LockNode *lock = new LockNode(C, tf);
3500 }
3501 #endif
3502
3503 return flock;
3504 }
3505
3506
3507 //------------------------------shared_unlock----------------------------------
3508 // Emit unlocking code.
3509 void GraphKit::shared_unlock(Node* box, Node* obj) {
3510 // bci is either a monitorenter bc or InvocationEntryBci
3511 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3512 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3513
3514 if( !GenerateSynchronizationCode )
3515 return;
3516 if (stopped()) { // Dead monitor?
3517 map()->pop_monitor(); // Kill monitor from debug info
3518 return;
3519 }
3520
3521 // Memory barrier to avoid floating things down past the locked region
3522 insert_mem_bar(Op_MemBarReleaseLock);
3523
3524 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3525 UnlockNode *unlock = new UnlockNode(C, tf);
3526 #ifdef ASSERT
3527 unlock->set_dbg_jvms(sync_jvms());
3528 #endif
3529 uint raw_idx = Compile::AliasIdxRaw;
3530 unlock->init_req( TypeFunc::Control, control() );
3531 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3532 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3533 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3534 unlock->init_req( TypeFunc::ReturnAdr, top() );
3535
3536 unlock->init_req(TypeFunc::Parms + 0, obj);
3537 unlock->init_req(TypeFunc::Parms + 1, box);
3538 unlock = _gvn.transform(unlock)->as_Unlock();
3539
3540 Node* mem = reset_memory();
3541
3542 // unlock has no side-effects, sets few values
3543 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3544
3545 // Kill monitor from debug info
3546 map()->pop_monitor( );
3547 }
3548
3549 //-------------------------------get_layout_helper-----------------------------
3550 // If the given klass is a constant or known to be an array,
3551 // fetch the constant layout helper value into constant_value
3552 // and return null. Otherwise, load the non-constant
3553 // layout helper value, and return the node which represents it.
3554 // This two-faced routine is useful because allocation sites
3555 // almost always feature constant types.
3556 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3557 const TypeKlassPtr* klass_t = _gvn.type(klass_node)->isa_klassptr();
3558 if (!StressReflectiveCode && klass_t != nullptr) {
3559 bool xklass = klass_t->klass_is_exact();
3560 if (xklass || (klass_t->isa_aryklassptr() && klass_t->is_aryklassptr()->elem() != Type::BOTTOM)) {
3561 jint lhelper;
3562 if (klass_t->isa_aryklassptr()) {
3563 BasicType elem = klass_t->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
3564 if (is_reference_type(elem, true)) {
3565 elem = T_OBJECT;
3566 }
3567 lhelper = Klass::array_layout_helper(elem);
3568 } else {
3569 lhelper = klass_t->is_instklassptr()->exact_klass()->layout_helper();
3570 }
3571 if (lhelper != Klass::_lh_neutral_value) {
3572 constant_value = lhelper;
3573 return (Node*) nullptr;
3574 }
3575 }
3576 }
3577 constant_value = Klass::_lh_neutral_value; // put in a known value
3578 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3579 return make_load(nullptr, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3580 }
3581
3582 // We just put in an allocate/initialize with a big raw-memory effect.
3583 // Hook selected additional alias categories on the initialization.
3584 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3585 MergeMemNode* init_in_merge,
3586 Node* init_out_raw) {
3587 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3588 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3589
3590 Node* prevmem = kit.memory(alias_idx);
3591 init_in_merge->set_memory_at(alias_idx, prevmem);
3592 kit.set_memory(init_out_raw, alias_idx);
3593 }
3594
3595 //---------------------------set_output_for_allocation-------------------------
3596 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3597 const TypeOopPtr* oop_type,
3598 bool deoptimize_on_exception) {
3599 int rawidx = Compile::AliasIdxRaw;
3600 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3601 add_safepoint_edges(alloc);
3602 Node* allocx = _gvn.transform(alloc);
3603 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3604 // create memory projection for i_o
3605 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3606 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3607
3608 // create a memory projection as for the normal control path
3609 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3610 set_memory(malloc, rawidx);
3611
3612 // a normal slow-call doesn't change i_o, but an allocation does
3613 // we create a separate i_o projection for the normal control path
3614 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3615 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3616
3617 // put in an initialization barrier
3618 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3619 rawoop)->as_Initialize();
3620 assert(alloc->initialization() == init, "2-way macro link must work");
3621 assert(init ->allocation() == alloc, "2-way macro link must work");
3622 {
3623 // Extract memory strands which may participate in the new object's
3624 // initialization, and source them from the new InitializeNode.
3625 // This will allow us to observe initializations when they occur,
3626 // and link them properly (as a group) to the InitializeNode.
3627 assert(init->in(InitializeNode::Memory) == malloc, "");
3628 MergeMemNode* minit_in = MergeMemNode::make(malloc);
3629 init->set_req(InitializeNode::Memory, minit_in);
3630 record_for_igvn(minit_in); // fold it up later, if possible
3631 Node* minit_out = memory(rawidx);
3632 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3633 // Add an edge in the MergeMem for the header fields so an access
3634 // to one of those has correct memory state
3635 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes())));
3636 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes())));
3637 if (oop_type->isa_aryptr()) {
3638 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3639 int elemidx = C->get_alias_index(telemref);
3640 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3641 } else if (oop_type->isa_instptr()) {
3642 ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass();
3643 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3644 ciField* field = ik->nonstatic_field_at(i);
3645 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
3646 continue; // do not bother to track really large numbers of fields
3647 // Find (or create) the alias category for this field:
3648 int fieldidx = C->alias_type(field)->index();
3649 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3650 }
3651 }
3652 }
3653
3654 // Cast raw oop to the real thing...
3655 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3656 javaoop = _gvn.transform(javaoop);
3657 C->set_recent_alloc(control(), javaoop);
3658 assert(just_allocated_object(control()) == javaoop, "just allocated");
3659
3660 #ifdef ASSERT
3661 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3672 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3673 }
3674 }
3675 #endif //ASSERT
3676
3677 return javaoop;
3678 }
3679
3680 //---------------------------new_instance--------------------------------------
3681 // This routine takes a klass_node which may be constant (for a static type)
3682 // or may be non-constant (for reflective code). It will work equally well
3683 // for either, and the graph will fold nicely if the optimizer later reduces
3684 // the type to a constant.
3685 // The optional arguments are for specialized use by intrinsics:
3686 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3687 // - If 'return_size_val', report the total object size to the caller.
3688 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3689 Node* GraphKit::new_instance(Node* klass_node,
3690 Node* extra_slow_test,
3691 Node* *return_size_val,
3692 bool deoptimize_on_exception) {
3693 // Compute size in doublewords
3694 // The size is always an integral number of doublewords, represented
3695 // as a positive bytewise size stored in the klass's layout_helper.
3696 // The layout_helper also encodes (in a low bit) the need for a slow path.
3697 jint layout_con = Klass::_lh_neutral_value;
3698 Node* layout_val = get_layout_helper(klass_node, layout_con);
3699 int layout_is_con = (layout_val == nullptr);
3700
3701 if (extra_slow_test == nullptr) extra_slow_test = intcon(0);
3702 // Generate the initial go-slow test. It's either ALWAYS (return a
3703 // Node for 1) or NEVER (return a null) or perhaps (in the reflective
3704 // case) a computed value derived from the layout_helper.
3705 Node* initial_slow_test = nullptr;
3706 if (layout_is_con) {
3707 assert(!StressReflectiveCode, "stress mode does not use these paths");
3708 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3709 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3710 } else { // reflective case
3711 // This reflective path is used by Unsafe.allocateInstance.
3712 // (It may be stress-tested by specifying StressReflectiveCode.)
3713 // Basically, we want to get into the VM is there's an illegal argument.
3714 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3715 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3716 if (extra_slow_test != intcon(0)) {
3717 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3718 }
3719 // (Macro-expander will further convert this to a Bool, if necessary.)
3730
3731 // Clear the low bits to extract layout_helper_size_in_bytes:
3732 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3733 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3734 size = _gvn.transform( new AndXNode(size, mask) );
3735 }
3736 if (return_size_val != nullptr) {
3737 (*return_size_val) = size;
3738 }
3739
3740 // This is a precise notnull oop of the klass.
3741 // (Actually, it need not be precise if this is a reflective allocation.)
3742 // It's what we cast the result to.
3743 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3744 if (!tklass) tklass = TypeInstKlassPtr::OBJECT;
3745 const TypeOopPtr* oop_type = tklass->as_instance_type();
3746
3747 // Now generate allocation code
3748
3749 // The entire memory state is needed for slow path of the allocation
3750 // since GC and deoptimization can happened.
3751 Node *mem = reset_memory();
3752 set_all_memory(mem); // Create new memory state
3753
3754 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3755 control(), mem, i_o(),
3756 size, klass_node,
3757 initial_slow_test);
3758
3759 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3760 }
3761
3762 //-------------------------------new_array-------------------------------------
3763 // helper for both newarray and anewarray
3764 // The 'length' parameter is (obviously) the length of the array.
3765 // The optional arguments are for specialized use by intrinsics:
3766 // - If 'return_size_val', report the non-padded array size (sum of header size
3767 // and array body) to the caller.
3768 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3769 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3770 Node* length, // number of array elements
3771 int nargs, // number of arguments to push back for uncommon trap
3772 Node* *return_size_val,
3773 bool deoptimize_on_exception) {
3774 jint layout_con = Klass::_lh_neutral_value;
3775 Node* layout_val = get_layout_helper(klass_node, layout_con);
3776 int layout_is_con = (layout_val == nullptr);
3777
3778 if (!layout_is_con && !StressReflectiveCode &&
3779 !too_many_traps(Deoptimization::Reason_class_check)) {
3780 // This is a reflective array creation site.
3781 // Optimistically assume that it is a subtype of Object[],
3782 // so that we can fold up all the address arithmetic.
3783 layout_con = Klass::array_layout_helper(T_OBJECT);
3784 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3785 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3786 { BuildCutout unless(this, bol_lh, PROB_MAX);
3787 inc_sp(nargs);
3788 uncommon_trap(Deoptimization::Reason_class_check,
3789 Deoptimization::Action_maybe_recompile);
3790 }
3791 layout_val = nullptr;
3792 layout_is_con = true;
3793 }
3794
3795 // Generate the initial go-slow test. Make sure we do not overflow
3796 // if length is huge (near 2Gig) or negative! We do not need
3797 // exact double-words here, just a close approximation of needed
3798 // double-words. We can't add any offset or rounding bits, lest we
3799 // take a size -1 of bytes and make it positive. Use an unsigned
3800 // compare, so negative sizes look hugely positive.
3801 int fast_size_limit = FastAllocateSizeLimit;
3802 if (layout_is_con) {
3803 assert(!StressReflectiveCode, "stress mode does not use these paths");
3804 // Increase the size limit if we have exact knowledge of array type.
3805 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3806 assert(fast_size_limit == 0 || count_leading_zeros(fast_size_limit) > static_cast<unsigned>(LogBytesPerLong - log2_esize),
3807 "fast_size_limit (%d) overflow when shifted left by %d", fast_size_limit, LogBytesPerLong - log2_esize);
3808 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3809 }
3810
3811 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3812 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3813
3814 // --- Size Computation ---
3815 // array_size = round_to_heap(array_header + (length << elem_shift));
3816 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
3817 // and align_to(x, y) == ((x + y-1) & ~(y-1))
3818 // The rounding mask is strength-reduced, if possible.
3819 int round_mask = MinObjAlignmentInBytes - 1;
3820 Node* header_size = nullptr;
3821 // (T_BYTE has the weakest alignment and size restrictions...)
3822 if (layout_is_con) {
3823 int hsize = Klass::layout_helper_header_size(layout_con);
3824 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3825 if ((round_mask & ~right_n_bits(eshift)) == 0)
3826 round_mask = 0; // strength-reduce it if it goes away completely
3827 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3828 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3829 assert(header_size_min <= hsize, "generic minimum is smallest");
3830 header_size = intcon(hsize);
3831 } else {
3832 Node* hss = intcon(Klass::_lh_header_size_shift);
3833 Node* hsm = intcon(Klass::_lh_header_size_mask);
3834 header_size = _gvn.transform(new URShiftINode(layout_val, hss));
3835 header_size = _gvn.transform(new AndINode(header_size, hsm));
3836 }
3837
3838 Node* elem_shift = nullptr;
3839 if (layout_is_con) {
3840 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3841 if (eshift != 0)
3842 elem_shift = intcon(eshift);
3843 } else {
3844 // There is no need to mask or shift this value.
3845 // The semantics of LShiftINode include an implicit mask to 0x1F.
3846 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3847 elem_shift = layout_val;
3896 }
3897 Node* non_rounded_size = _gvn.transform(new AddXNode(headerx, abody));
3898
3899 if (return_size_val != nullptr) {
3900 // This is the size
3901 (*return_size_val) = non_rounded_size;
3902 }
3903
3904 Node* size = non_rounded_size;
3905 if (round_mask != 0) {
3906 Node* mask1 = MakeConX(round_mask);
3907 size = _gvn.transform(new AddXNode(size, mask1));
3908 Node* mask2 = MakeConX(~round_mask);
3909 size = _gvn.transform(new AndXNode(size, mask2));
3910 }
3911 // else if round_mask == 0, the size computation is self-rounding
3912
3913 // Now generate allocation code
3914
3915 // The entire memory state is needed for slow path of the allocation
3916 // since GC and deoptimization can happened.
3917 Node *mem = reset_memory();
3918 set_all_memory(mem); // Create new memory state
3919
3920 if (initial_slow_test->is_Bool()) {
3921 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3922 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3923 }
3924
3925 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3926 Node* valid_length_test = _gvn.intcon(1);
3927 if (ary_type->isa_aryptr()) {
3928 BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type();
3929 jint max = TypeAryPtr::max_array_length(bt);
3930 Node* valid_length_cmp = _gvn.transform(new CmpUNode(length, intcon(max)));
3931 valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le));
3932 }
3933
3934 // Create the AllocateArrayNode and its result projections
3935 AllocateArrayNode* alloc
3936 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3937 control(), mem, i_o(),
3938 size, klass_node,
3939 initial_slow_test,
3940 length, valid_length_test);
3941
3942 // Cast to correct type. Note that the klass_node may be constant or not,
3943 // and in the latter case the actual array type will be inexact also.
3944 // (This happens via a non-constant argument to inline_native_newArray.)
3945 // In any case, the value of klass_node provides the desired array type.
3946 const TypeInt* length_type = _gvn.find_int_type(length);
3947 if (ary_type->isa_aryptr() && length_type != nullptr) {
3948 // Try to get a better type than POS for the size
3949 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3950 }
3951
3952 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3953
3954 array_ideal_length(alloc, ary_type, true);
3955 return javaoop;
3956 }
3957
3958 // The following "Ideal_foo" functions are placed here because they recognize
3959 // the graph shapes created by the functions immediately above.
3960
3961 //---------------------------Ideal_allocation----------------------------------
4069 set_all_memory(ideal.merged_memory());
4070 set_i_o(ideal.i_o());
4071 set_control(ideal.ctrl());
4072 }
4073
4074 void GraphKit::final_sync(IdealKit& ideal) {
4075 // Final sync IdealKit and graphKit.
4076 sync_kit(ideal);
4077 }
4078
4079 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4080 Node* len = load_array_length(load_String_value(str, set_ctrl));
4081 Node* coder = load_String_coder(str, set_ctrl);
4082 // Divide length by 2 if coder is UTF16
4083 return _gvn.transform(new RShiftINode(len, coder));
4084 }
4085
4086 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4087 int value_offset = java_lang_String::value_offset();
4088 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4089 false, nullptr, 0);
4090 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4091 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4092 TypeAry::make(TypeInt::BYTE, TypeInt::POS),
4093 ciTypeArrayKlass::make(T_BYTE), true, 0);
4094 Node* p = basic_plus_adr(str, str, value_offset);
4095 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4096 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4097 return load;
4098 }
4099
4100 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4101 if (!CompactStrings) {
4102 return intcon(java_lang_String::CODER_UTF16);
4103 }
4104 int coder_offset = java_lang_String::coder_offset();
4105 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4106 false, nullptr, 0);
4107 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4108
4109 Node* p = basic_plus_adr(str, str, coder_offset);
4110 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4111 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4112 return load;
4113 }
4114
4115 void GraphKit::store_String_value(Node* str, Node* value) {
4116 int value_offset = java_lang_String::value_offset();
4117 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4118 false, nullptr, 0);
4119 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4120
4121 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4122 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4123 }
4124
4125 void GraphKit::store_String_coder(Node* str, Node* value) {
4126 int coder_offset = java_lang_String::coder_offset();
4127 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4128 false, nullptr, 0);
4129 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4130
4131 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4132 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4133 }
4134
4135 // Capture src and dst memory state with a MergeMemNode
4136 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4137 if (src_type == dst_type) {
4138 // Types are equal, we don't need a MergeMemNode
4139 return memory(src_type);
4140 }
4141 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4142 record_for_igvn(merge); // fold it up later, if possible
4143 int src_idx = C->get_alias_index(src_type);
4144 int dst_idx = C->get_alias_index(dst_type);
4145 merge->set_memory_at(src_idx, memory(src_idx));
4146 merge->set_memory_at(dst_idx, memory(dst_idx));
4147 return merge;
4148 }
4221 i_char->init_req(2, AddI(i_char, intcon(2)));
4222
4223 set_control(IfFalse(iff));
4224 set_memory(st, TypeAryPtr::BYTES);
4225 }
4226
4227 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4228 if (!field->is_constant()) {
4229 return nullptr; // Field not marked as constant.
4230 }
4231 ciInstance* holder = nullptr;
4232 if (!field->is_static()) {
4233 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4234 if (const_oop != nullptr && const_oop->is_instance()) {
4235 holder = const_oop->as_instance();
4236 }
4237 }
4238 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4239 /*is_unsigned_load=*/false);
4240 if (con_type != nullptr) {
4241 return makecon(con_type);
4242 }
4243 return nullptr;
4244 }
4245
4246 Node* GraphKit::maybe_narrow_object_type(Node* obj, ciKlass* type) {
4247 const TypeOopPtr* obj_type = obj->bottom_type()->isa_oopptr();
4248 const TypeOopPtr* sig_type = TypeOopPtr::make_from_klass(type);
4249 if (obj_type != nullptr && sig_type->is_loaded() && !obj_type->higher_equal(sig_type)) {
4250 const Type* narrow_obj_type = obj_type->filter_speculative(sig_type); // keep speculative part
4251 Node* casted_obj = gvn().transform(new CheckCastPPNode(control(), obj, narrow_obj_type));
4252 return casted_obj;
4253 }
4254 return obj;
4255 }
|
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "asm/register.hpp"
26 #include "ci/ciFlatArrayKlass.hpp"
27 #include "ci/ciInlineKlass.hpp"
28 #include "ci/ciObjArray.hpp"
29 #include "ci/ciUtilities.hpp"
30 #include "classfile/javaClasses.hpp"
31 #include "compiler/compileLog.hpp"
32 #include "gc/shared/barrierSet.hpp"
33 #include "gc/shared/c2/barrierSetC2.hpp"
34 #include "interpreter/interpreter.hpp"
35 #include "memory/resourceArea.hpp"
36 #include "oops/flatArrayKlass.hpp"
37 #include "opto/addnode.hpp"
38 #include "opto/castnode.hpp"
39 #include "opto/convertnode.hpp"
40 #include "opto/graphKit.hpp"
41 #include "opto/idealKit.hpp"
42 #include "opto/inlinetypenode.hpp"
43 #include "opto/intrinsicnode.hpp"
44 #include "opto/locknode.hpp"
45 #include "opto/machnode.hpp"
46 #include "opto/narrowptrnode.hpp"
47 #include "opto/opaquenode.hpp"
48 #include "opto/parse.hpp"
49 #include "opto/rootnode.hpp"
50 #include "opto/runtime.hpp"
51 #include "opto/subtypenode.hpp"
52 #include "runtime/deoptimization.hpp"
53 #include "runtime/sharedRuntime.hpp"
54 #include "runtime/stubRoutines.hpp"
55 #include "utilities/bitMap.inline.hpp"
56 #include "utilities/growableArray.hpp"
57 #include "utilities/powerOfTwo.hpp"
58
59 //----------------------------GraphKit-----------------------------------------
60 // Main utility constructor.
61 GraphKit::GraphKit(JVMState* jvms, PhaseGVN* gvn)
62 : Phase(Phase::Parser),
63 _env(C->env()),
64 _gvn((gvn != nullptr) ? *gvn : *C->initial_gvn()),
65 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
66 {
67 assert(gvn == nullptr || !gvn->is_IterGVN() || gvn->is_IterGVN()->delay_transform(), "delay transform should be enabled");
68 _exceptions = jvms->map()->next_exception();
69 if (_exceptions != nullptr) jvms->map()->set_next_exception(nullptr);
70 set_jvms(jvms);
71 #ifdef ASSERT
72 if (_gvn.is_IterGVN() != nullptr) {
73 assert(_gvn.is_IterGVN()->delay_transform(), "Transformation must be delayed if IterGVN is used");
74 // Save the initial size of _for_igvn worklist for verification (see ~GraphKit)
75 _worklist_size = _gvn.C->igvn_worklist()->size();
76 }
77 #endif
78 }
79
80 // Private constructor for parser.
81 GraphKit::GraphKit()
82 : Phase(Phase::Parser),
83 _env(C->env()),
84 _gvn(*C->initial_gvn()),
85 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
86 {
87 _exceptions = nullptr;
88 set_map(nullptr);
89 DEBUG_ONLY(_sp = -99);
90 DEBUG_ONLY(set_bci(-99));
91 }
92
93
94
95 //---------------------------clean_stack---------------------------------------
96 // Clear away rubbish from the stack area of the JVM state.
97 // This destroys any arguments that may be waiting on the stack.
342 }
343 static inline void add_one_req(Node* dstphi, Node* src) {
344 assert(is_hidden_merge(dstphi), "must be a special merge node");
345 assert(!is_hidden_merge(src), "must not be a special merge node");
346 dstphi->add_req(src);
347 }
348
349 //-----------------------combine_exception_states------------------------------
350 // This helper function combines exception states by building phis on a
351 // specially marked state-merging region. These regions and phis are
352 // untransformed, and can build up gradually. The region is marked by
353 // having a control input of its exception map, rather than null. Such
354 // regions do not appear except in this function, and in use_exception_state.
355 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
356 if (failing_internal()) {
357 return; // dying anyway...
358 }
359 JVMState* ex_jvms = ex_map->_jvms;
360 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
361 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
362 // TODO 8325632 Re-enable
363 // assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
364 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
365 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
366 assert(ex_map->req() == phi_map->req(), "matching maps");
367 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
368 Node* hidden_merge_mark = root();
369 Node* region = phi_map->control();
370 MergeMemNode* phi_mem = phi_map->merged_memory();
371 MergeMemNode* ex_mem = ex_map->merged_memory();
372 if (region->in(0) != hidden_merge_mark) {
373 // The control input is not (yet) a specially-marked region in phi_map.
374 // Make it so, and build some phis.
375 region = new RegionNode(2);
376 _gvn.set_type(region, Type::CONTROL);
377 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
378 region->init_req(1, phi_map->control());
379 phi_map->set_control(region);
380 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
381 record_for_igvn(io_phi);
382 _gvn.set_type(io_phi, Type::ABIO);
383 phi_map->set_i_o(io_phi);
871 if (PrintMiscellaneous && (Verbose || WizardMode)) {
872 tty->print_cr("Zombie local %d: ", local);
873 jvms->dump();
874 }
875 return false;
876 }
877 }
878 }
879 return true;
880 }
881
882 #endif //ASSERT
883
884 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens.
885 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
886 ciMethod* cur_method = jvms->method();
887 int cur_bci = jvms->bci();
888 if (cur_method != nullptr && cur_bci != InvocationEntryBci) {
889 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
890 return Interpreter::bytecode_should_reexecute(code) ||
891 (is_anewarray && (code == Bytecodes::_multianewarray));
892 // Reexecute _multianewarray bytecode which was replaced with
893 // sequence of [a]newarray. See Parse::do_multianewarray().
894 //
895 // Note: interpreter should not have it set since this optimization
896 // is limited by dimensions and guarded by flag so in some cases
897 // multianewarray() runtime calls will be generated and
898 // the bytecode should not be reexecutes (stack will not be reset).
899 } else {
900 return false;
901 }
902 }
903
904 // Helper function for adding JVMState and debug information to node
905 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
906 // Add the safepoint edges to the call (or other safepoint).
907
908 // Make sure dead locals are set to top. This
909 // should help register allocation time and cut down on the size
910 // of the deoptimization information.
911 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
962 }
963
964 // Presize the call:
965 DEBUG_ONLY(uint non_debug_edges = call->req());
966 call->add_req_batch(top(), youngest_jvms->debug_depth());
967 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
968
969 // Set up edges so that the call looks like this:
970 // Call [state:] ctl io mem fptr retadr
971 // [parms:] parm0 ... parmN
972 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
973 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
974 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
975 // Note that caller debug info precedes callee debug info.
976
977 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
978 uint debug_ptr = call->req();
979
980 // Loop over the map input edges associated with jvms, add them
981 // to the call node, & reset all offsets to match call node array.
982
983 JVMState* callee_jvms = nullptr;
984 for (JVMState* in_jvms = youngest_jvms; in_jvms != nullptr; ) {
985 uint debug_end = debug_ptr;
986 uint debug_start = debug_ptr - in_jvms->debug_size();
987 debug_ptr = debug_start; // back up the ptr
988
989 uint p = debug_start; // walks forward in [debug_start, debug_end)
990 uint j, k, l;
991 SafePointNode* in_map = in_jvms->map();
992 out_jvms->set_map(call);
993
994 if (can_prune_locals) {
995 assert(in_jvms->method() == out_jvms->method(), "sanity");
996 // If the current throw can reach an exception handler in this JVMS,
997 // then we must keep everything live that can reach that handler.
998 // As a quick and dirty approximation, we look for any handlers at all.
999 if (in_jvms->method()->has_exception_handlers()) {
1000 can_prune_locals = false;
1001 }
1002 }
1003
1004 // Add the Locals
1005 k = in_jvms->locoff();
1006 l = in_jvms->loc_size();
1007 out_jvms->set_locoff(p);
1008 if (!can_prune_locals) {
1009 for (j = 0; j < l; j++) {
1010 call->set_req(p++, in_map->in(k + j));
1011 }
1012 } else {
1013 p += l; // already set to top above by add_req_batch
1014 }
1015
1016 // Add the Expression Stack
1017 k = in_jvms->stkoff();
1018 l = in_jvms->sp();
1019 out_jvms->set_stkoff(p);
1020 if (!can_prune_locals) {
1021 for (j = 0; j < l; j++) {
1022 call->set_req(p++, in_map->in(k + j));
1023 }
1024 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
1025 // Divide stack into {S0,...,S1}, where S0 is set to top.
1026 uint s1 = stack_slots_not_pruned;
1027 stack_slots_not_pruned = 0; // for next iteration
1028 if (s1 > l) s1 = l;
1029 uint s0 = l - s1;
1030 p += s0; // skip the tops preinstalled by add_req_batch
1031 for (j = s0; j < l; j++)
1032 call->set_req(p++, in_map->in(k+j));
1033 } else {
1034 p += l; // already set to top above by add_req_batch
1035 }
1036
1037 // Add the Monitors
1038 k = in_jvms->monoff();
1039 l = in_jvms->mon_size();
1040 out_jvms->set_monoff(p);
1041 for (j = 0; j < l; j++)
1042 call->set_req(p++, in_map->in(k+j));
1043
1044 // Copy any scalar object fields.
1045 k = in_jvms->scloff();
1046 l = in_jvms->scl_size();
1047 out_jvms->set_scloff(p);
1048 for (j = 0; j < l; j++)
1049 call->set_req(p++, in_map->in(k+j));
1050
1051 // Finish the new jvms.
1052 out_jvms->set_endoff(p);
1053
1054 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
1055 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
1056 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
1057 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
1058 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
1059 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
1060
1061 // Update the two tail pointers in parallel.
1062 callee_jvms = out_jvms;
1063 out_jvms = out_jvms->caller();
1064 in_jvms = in_jvms->caller();
1065 }
1066
1067 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
1068
1069 // Test the correctness of JVMState::debug_xxx accessors:
1070 assert(call->jvms()->debug_start() == non_debug_edges, "");
1071 assert(call->jvms()->debug_end() == call->req(), "");
1072 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1073 }
1074
1075 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1076 Bytecodes::Code code = java_bc();
1077 if (code == Bytecodes::_wide) {
1078 code = method()->java_code_at_bci(bci() + 1);
1079 }
1080
1081 if (code != Bytecodes::_illegal) {
1082 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1218 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1219 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1220 return _gvn.transform( new AndLNode(conv, mask) );
1221 }
1222
1223 Node* GraphKit::ConvL2I(Node* offset) {
1224 // short-circuit a common case
1225 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1226 if (offset_con != (jlong)Type::OffsetBot) {
1227 return intcon((int) offset_con);
1228 }
1229 return _gvn.transform( new ConvL2INode(offset));
1230 }
1231
1232 //-------------------------load_object_klass-----------------------------------
1233 Node* GraphKit::load_object_klass(Node* obj) {
1234 // Special-case a fresh allocation to avoid building nodes:
1235 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1236 if (akls != nullptr) return akls;
1237 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1238 return _gvn.transform(LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
1239 }
1240
1241 //-------------------------load_array_length-----------------------------------
1242 Node* GraphKit::load_array_length(Node* array) {
1243 // Special-case a fresh allocation to avoid building nodes:
1244 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array);
1245 Node *alen;
1246 if (alloc == nullptr) {
1247 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1248 alen = _gvn.transform( new LoadRangeNode(nullptr, immutable_memory(), r_adr, TypeInt::POS));
1249 } else {
1250 alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false);
1251 }
1252 return alen;
1253 }
1254
1255 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc,
1256 const TypeOopPtr* oop_type,
1257 bool replace_length_in_map) {
1258 Node* length = alloc->Ideal_length();
1267 replace_in_map(length, ccast);
1268 }
1269 return ccast;
1270 }
1271 }
1272 return length;
1273 }
1274
1275 //------------------------------do_null_check----------------------------------
1276 // Helper function to do a null pointer check. Returned value is
1277 // the incoming address with null casted away. You are allowed to use the
1278 // not-null value only if you are control dependent on the test.
1279 #ifndef PRODUCT
1280 extern uint explicit_null_checks_inserted,
1281 explicit_null_checks_elided;
1282 #endif
1283 Node* GraphKit::null_check_common(Node* value, BasicType type,
1284 // optional arguments for variations:
1285 bool assert_null,
1286 Node* *null_control,
1287 bool speculative,
1288 bool null_marker_check) {
1289 assert(!assert_null || null_control == nullptr, "not both at once");
1290 if (stopped()) return top();
1291 NOT_PRODUCT(explicit_null_checks_inserted++);
1292
1293 if (value->is_InlineType()) {
1294 // Null checking a scalarized but nullable inline type. Check the null marker
1295 // input instead of the oop input to avoid keeping buffer allocations alive.
1296 InlineTypeNode* vtptr = value->as_InlineType();
1297 while (vtptr->get_oop()->is_InlineType()) {
1298 vtptr = vtptr->get_oop()->as_InlineType();
1299 }
1300 null_check_common(vtptr->get_null_marker(), T_INT, assert_null, null_control, speculative, true);
1301 if (stopped()) {
1302 return top();
1303 }
1304 if (assert_null) {
1305 // TODO 8284443 Scalarize here (this currently leads to compilation bailouts)
1306 // vtptr = InlineTypeNode::make_null(_gvn, vtptr->type()->inline_klass());
1307 // replace_in_map(value, vtptr);
1308 // return vtptr;
1309 replace_in_map(value, null());
1310 return null();
1311 }
1312 bool do_replace_in_map = (null_control == nullptr || (*null_control) == top());
1313 return cast_not_null(value, do_replace_in_map);
1314 }
1315
1316 // Construct null check
1317 Node *chk = nullptr;
1318 switch(type) {
1319 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1320 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1321 case T_ARRAY : // fall through
1322 type = T_OBJECT; // simplify further tests
1323 case T_OBJECT : {
1324 const Type *t = _gvn.type( value );
1325
1326 const TypeOopPtr* tp = t->isa_oopptr();
1327 if (tp != nullptr && !tp->is_loaded()
1328 // Only for do_null_check, not any of its siblings:
1329 && !assert_null && null_control == nullptr) {
1330 // Usually, any field access or invocation on an unloaded oop type
1331 // will simply fail to link, since the statically linked class is
1332 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1333 // the static class is loaded but the sharper oop type is not.
1334 // Rather than checking for this obscure case in lots of places,
1335 // we simply observe that a null check on an unloaded class
1399 }
1400 Node *oldcontrol = control();
1401 set_control(cfg);
1402 Node *res = cast_not_null(value);
1403 set_control(oldcontrol);
1404 NOT_PRODUCT(explicit_null_checks_elided++);
1405 return res;
1406 }
1407 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1408 if (cfg == nullptr) break; // Quit at region nodes
1409 depth++;
1410 }
1411 }
1412
1413 //-----------
1414 // Branch to failure if null
1415 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1416 Deoptimization::DeoptReason reason;
1417 if (assert_null) {
1418 reason = Deoptimization::reason_null_assert(speculative);
1419 } else if (type == T_OBJECT || null_marker_check) {
1420 reason = Deoptimization::reason_null_check(speculative);
1421 } else {
1422 reason = Deoptimization::Reason_div0_check;
1423 }
1424 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1425 // ciMethodData::has_trap_at will return a conservative -1 if any
1426 // must-be-null assertion has failed. This could cause performance
1427 // problems for a method after its first do_null_assert failure.
1428 // Consider using 'Reason_class_check' instead?
1429
1430 // To cause an implicit null check, we set the not-null probability
1431 // to the maximum (PROB_MAX). For an explicit check the probability
1432 // is set to a smaller value.
1433 if (null_control != nullptr || too_many_traps(reason)) {
1434 // probability is less likely
1435 ok_prob = PROB_LIKELY_MAG(3);
1436 } else if (!assert_null &&
1437 (ImplicitNullCheckThreshold > 0) &&
1438 method() != nullptr &&
1439 (method()->method_data()->trap_count(reason)
1473 }
1474
1475 if (assert_null) {
1476 // Cast obj to null on this path.
1477 replace_in_map(value, zerocon(type));
1478 return zerocon(type);
1479 }
1480
1481 // Cast obj to not-null on this path, if there is no null_control.
1482 // (If there is a null_control, a non-null value may come back to haunt us.)
1483 if (type == T_OBJECT) {
1484 Node* cast = cast_not_null(value, false);
1485 if (null_control == nullptr || (*null_control) == top())
1486 replace_in_map(value, cast);
1487 value = cast;
1488 }
1489
1490 return value;
1491 }
1492
1493 //------------------------------cast_not_null----------------------------------
1494 // Cast obj to not-null on this path
1495 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1496 if (obj->is_InlineType()) {
1497 Node* vt = obj->isa_InlineType()->clone_if_required(&gvn(), map(), do_replace_in_map);
1498 vt->as_InlineType()->set_null_marker(_gvn);
1499 vt = _gvn.transform(vt);
1500 if (do_replace_in_map) {
1501 replace_in_map(obj, vt);
1502 }
1503 return vt;
1504 }
1505 const Type *t = _gvn.type(obj);
1506 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1507 // Object is already not-null?
1508 if( t == t_not_null ) return obj;
1509
1510 Node* cast = new CastPPNode(control(), obj,t_not_null);
1511 cast = _gvn.transform( cast );
1512
1513 // Scan for instances of 'obj' in the current JVM mapping.
1514 // These instances are known to be not-null after the test.
1515 if (do_replace_in_map)
1516 replace_in_map(obj, cast);
1517
1518 return cast; // Return casted value
1519 }
1520
1521 Node* GraphKit::cast_to_non_larval(Node* obj) {
1522 const Type* obj_type = gvn().type(obj);
1523 if (obj->is_InlineType() || !obj_type->is_inlinetypeptr()) {
1524 return obj;
1525 }
1526
1527 Node* new_obj = InlineTypeNode::make_from_oop(this, obj, obj_type->inline_klass());
1528 replace_in_map(obj, new_obj);
1529 return new_obj;
1530 }
1531
1532 // Sometimes in intrinsics, we implicitly know an object is not null
1533 // (there's no actual null check) so we can cast it to not null. In
1534 // the course of optimizations, the input to the cast can become null.
1535 // In that case that data path will die and we need the control path
1536 // to become dead as well to keep the graph consistent. So we have to
1537 // add a check for null for which one branch can't be taken. It uses
1538 // an OpaqueNotNull node that will cause the check to be removed after loop
1539 // opts so the test goes away and the compiled code doesn't execute a
1540 // useless check.
1541 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1542 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) {
1543 return value;
1544 }
1545 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1546 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1547 Node* opaq = _gvn.transform(new OpaqueNotNullNode(C, tst));
1548 IfNode* iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1549 _gvn.set_type(iff, iff->Value(&_gvn));
1550 if (!tst->is_Con()) {
1551 record_for_igvn(iff);
1624 // These are layered on top of the factory methods in LoadNode and StoreNode,
1625 // and integrate with the parser's memory state and _gvn engine.
1626 //
1627
1628 // factory methods in "int adr_idx"
1629 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1630 MemNode::MemOrd mo,
1631 LoadNode::ControlDependency control_dependency,
1632 bool require_atomic_access,
1633 bool unaligned,
1634 bool mismatched,
1635 bool unsafe,
1636 uint8_t barrier_data) {
1637 int adr_idx = C->get_alias_index(_gvn.type(adr)->isa_ptr());
1638 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1639 const TypePtr* adr_type = nullptr; // debug-mode-only argument
1640 DEBUG_ONLY(adr_type = C->get_adr_type(adr_idx));
1641 Node* mem = memory(adr_idx);
1642 Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data);
1643 ld = _gvn.transform(ld);
1644
1645 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1646 // Improve graph before escape analysis and boxing elimination.
1647 record_for_igvn(ld);
1648 if (ld->is_DecodeN()) {
1649 // Also record the actual load (LoadN) in case ld is DecodeN. In some
1650 // rare corner cases, ld->in(1) can be something other than LoadN (e.g.,
1651 // a Phi). Recording such cases is still perfectly sound, but may be
1652 // unnecessary and result in some minor IGVN overhead.
1653 record_for_igvn(ld->in(1));
1654 }
1655 }
1656 return ld;
1657 }
1658
1659 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1660 MemNode::MemOrd mo,
1661 bool require_atomic_access,
1662 bool unaligned,
1663 bool mismatched,
1664 bool unsafe,
1678 if (unsafe) {
1679 st->as_Store()->set_unsafe_access();
1680 }
1681 st->as_Store()->set_barrier_data(barrier_data);
1682 st = _gvn.transform(st);
1683 set_memory(st, adr_idx);
1684 // Back-to-back stores can only remove intermediate store with DU info
1685 // so push on worklist for optimizer.
1686 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1687 record_for_igvn(st);
1688
1689 return st;
1690 }
1691
1692 Node* GraphKit::access_store_at(Node* obj,
1693 Node* adr,
1694 const TypePtr* adr_type,
1695 Node* val,
1696 const Type* val_type,
1697 BasicType bt,
1698 DecoratorSet decorators,
1699 bool safe_for_replace,
1700 const InlineTypeNode* vt) {
1701 // Transformation of a value which could be null pointer (CastPP #null)
1702 // could be delayed during Parse (for example, in adjust_map_after_if()).
1703 // Execute transformation here to avoid barrier generation in such case.
1704 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1705 val = _gvn.makecon(TypePtr::NULL_PTR);
1706 }
1707
1708 if (stopped()) {
1709 return top(); // Dead path ?
1710 }
1711
1712 assert(val != nullptr, "not dead path");
1713 if (val->is_InlineType()) {
1714 // Store to non-flat field. Buffer the inline type and make sure
1715 // the store is re-executed if the allocation triggers deoptimization.
1716 PreserveReexecuteState preexecs(this);
1717 jvms()->set_should_reexecute(true);
1718 val = val->as_InlineType()->buffer(this, safe_for_replace);
1719 }
1720
1721 C2AccessValuePtr addr(adr, adr_type);
1722 C2AccessValue value(val, val_type);
1723 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr, nullptr, vt);
1724 if (access.is_raw()) {
1725 return _barrier_set->BarrierSetC2::store_at(access, value);
1726 } else {
1727 return _barrier_set->store_at(access, value);
1728 }
1729 }
1730
1731 Node* GraphKit::access_load_at(Node* obj, // containing obj
1732 Node* adr, // actual address to store val at
1733 const TypePtr* adr_type,
1734 const Type* val_type,
1735 BasicType bt,
1736 DecoratorSet decorators,
1737 Node* ctl) {
1738 if (stopped()) {
1739 return top(); // Dead path ?
1740 }
1741
1742 C2AccessValuePtr addr(adr, adr_type);
1743 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr, ctl);
1744 if (access.is_raw()) {
1745 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1746 } else {
1747 return _barrier_set->load_at(access, val_type);
1748 }
1749 }
1750
1751 Node* GraphKit::access_load(Node* adr, // actual address to load val at
1752 const Type* val_type,
1753 BasicType bt,
1754 DecoratorSet decorators) {
1755 if (stopped()) {
1756 return top(); // Dead path ?
1757 }
1758
1759 C2AccessValuePtr addr(adr, adr->bottom_type()->is_ptr());
1760 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, nullptr, addr);
1761 if (access.is_raw()) {
1762 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1763 } else {
1828 Node* new_val,
1829 const Type* value_type,
1830 BasicType bt,
1831 DecoratorSet decorators) {
1832 C2AccessValuePtr addr(adr, adr_type);
1833 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1834 if (access.is_raw()) {
1835 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1836 } else {
1837 return _barrier_set->atomic_add_at(access, new_val, value_type);
1838 }
1839 }
1840
1841 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1842 return _barrier_set->clone(this, src, dst, size, is_array);
1843 }
1844
1845 //-------------------------array_element_address-------------------------
1846 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1847 const TypeInt* sizetype, Node* ctrl) {
1848 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr();
1849 uint shift;
1850 if (arytype->is_flat() && arytype->klass_is_exact()) {
1851 // We can only determine the flat array layout statically if the klass is exact. Otherwise, we could have different
1852 // value classes at runtime with a potentially different layout. The caller needs to fall back to call
1853 // load/store_unknown_inline_Type() at runtime. We could return a sentinel node for the non-exact case but that
1854 // might mess with other GVN transformations in between. Thus, we just continue in the else branch normally, even
1855 // though we don't need the address node in this case and throw it away again.
1856 shift = arytype->flat_log_elem_size();
1857 } else {
1858 shift = exact_log2(type2aelembytes(elembt));
1859 }
1860 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1861
1862 // short-circuit a common case (saves lots of confusing waste motion)
1863 jint idx_con = find_int_con(idx, -1);
1864 if (idx_con >= 0) {
1865 intptr_t offset = header + ((intptr_t)idx_con << shift);
1866 return basic_plus_adr(ary, offset);
1867 }
1868
1869 // must be correct type for alignment purposes
1870 Node* base = basic_plus_adr(ary, header);
1871 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1872 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1873 return basic_plus_adr(ary, base, scale);
1874 }
1875
1876 Node* GraphKit::cast_to_flat_array(Node* array, ciInlineKlass* vk, bool is_null_free, bool is_not_null_free, bool is_atomic) {
1877 assert(vk->maybe_flat_in_array(), "element of type %s cannot be flat in array", vk->name()->as_utf8());
1878 if (!vk->has_nullable_atomic_layout()) {
1879 // Element does not have a nullable flat layout, cannot be nullable
1880 is_null_free = true;
1881 }
1882 if (!vk->has_atomic_layout() && !vk->has_non_atomic_layout()) {
1883 // Element does not have a null-free flat layout, cannot be null-free
1884 is_not_null_free = true;
1885 }
1886 if (is_null_free) {
1887 // TODO 8350865 Impossible type
1888 is_not_null_free = false;
1889 }
1890
1891 bool is_exact = is_null_free || is_not_null_free;
1892 ciArrayKlass* array_klass = ciArrayKlass::make(vk, is_null_free, is_atomic, true);
1893 assert(array_klass->is_elem_null_free() == is_null_free, "inconsistency");
1894 assert(array_klass->is_elem_atomic() == is_atomic, "inconsistency");
1895 const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr();
1896 arytype = arytype->cast_to_exactness(is_exact);
1897 arytype = arytype->cast_to_not_null_free(is_not_null_free);
1898 assert(arytype->is_null_free() == is_null_free, "inconsistency");
1899 assert(arytype->is_not_null_free() == is_not_null_free, "inconsistency");
1900 assert(arytype->is_atomic() == is_atomic, "inconsistency");
1901 return _gvn.transform(new CastPPNode(control(), array, arytype, ConstraintCastNode::StrongDependency));
1902 }
1903
1904 //-------------------------load_array_element-------------------------
1905 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) {
1906 const Type* elemtype = arytype->elem();
1907 BasicType elembt = elemtype->array_element_basic_type();
1908 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1909 if (elembt == T_NARROWOOP) {
1910 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1911 }
1912 Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt,
1913 IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0));
1914 return ld;
1915 }
1916
1917 //-------------------------set_arguments_for_java_call-------------------------
1918 // Arguments (pre-popped from the stack) are taken from the JVMS.
1919 void GraphKit::set_arguments_for_java_call(CallJavaNode* call, bool is_late_inline) {
1920 PreserveReexecuteState preexecs(this);
1921 if (EnableValhalla) {
1922 // Make sure the call is "re-executed", if buffering of inline type arguments triggers deoptimization.
1923 // At this point, the call hasn't been executed yet, so we will only ever execute the call once.
1924 jvms()->set_should_reexecute(true);
1925 int arg_size = method()->get_declared_signature_at_bci(bci())->arg_size_for_bc(java_bc());
1926 inc_sp(arg_size);
1927 }
1928 // Add the call arguments
1929 const TypeTuple* domain = call->tf()->domain_sig();
1930 uint nargs = domain->cnt();
1931 int arg_num = 0;
1932 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) {
1933 Node* arg = argument(i-TypeFunc::Parms);
1934 const Type* t = domain->field_at(i);
1935 // TODO 8284443 A static call to a mismatched method should still be scalarized
1936 if (t->is_inlinetypeptr() && !call->method()->get_Method()->mismatch() && call->method()->is_scalarized_arg(arg_num)) {
1937 // We don't pass inline type arguments by reference but instead pass each field of the inline type
1938 if (!arg->is_InlineType()) {
1939 assert(_gvn.type(arg)->is_zero_type() && !t->inline_klass()->is_null_free(), "Unexpected argument type");
1940 arg = InlineTypeNode::make_from_oop(this, arg, t->inline_klass());
1941 }
1942 InlineTypeNode* vt = arg->as_InlineType();
1943 vt->pass_fields(this, call, idx, true, !t->maybe_null());
1944 // If an inline type argument is passed as fields, attach the Method* to the call site
1945 // to be able to access the extended signature later via attached_method_before_pc().
1946 // For example, see CompiledMethod::preserve_callee_argument_oops().
1947 call->set_override_symbolic_info(true);
1948 // Register an evol dependency on the callee method to make sure that this method is deoptimized and
1949 // re-compiled with a non-scalarized calling convention if the callee method is later marked as mismatched.
1950 C->dependencies()->assert_evol_method(call->method());
1951 arg_num++;
1952 continue;
1953 } else if (arg->is_InlineType()) {
1954 // Pass inline type argument via oop to callee
1955 arg = arg->as_InlineType()->buffer(this, true);
1956 }
1957 if (t != Type::HALF) {
1958 arg_num++;
1959 }
1960 call->init_req(idx++, arg);
1961 }
1962 }
1963
1964 //---------------------------set_edges_for_java_call---------------------------
1965 // Connect a newly created call into the current JVMS.
1966 // A return value node (if any) is returned from set_edges_for_java_call.
1967 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1968
1969 // Add the predefined inputs:
1970 call->init_req( TypeFunc::Control, control() );
1971 call->init_req( TypeFunc::I_O , i_o() );
1972 call->init_req( TypeFunc::Memory , reset_memory() );
1973 call->init_req( TypeFunc::FramePtr, frameptr() );
1974 call->init_req( TypeFunc::ReturnAdr, top() );
1975
1976 add_safepoint_edges(call, must_throw);
1977
1978 Node* xcall = _gvn.transform(call);
1979
1980 if (xcall == top()) {
1981 set_control(top());
1982 return;
1983 }
1984 assert(xcall == call, "call identity is stable");
1985
1986 // Re-use the current map to produce the result.
1987
1988 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1989 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1990 set_all_memory_call(xcall, separate_io_proj);
1991
1992 //return xcall; // no need, caller already has it
1993 }
1994
1995 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
1996 if (stopped()) return top(); // maybe the call folded up?
1997
1998 // Note: Since any out-of-line call can produce an exception,
1999 // we always insert an I_O projection from the call into the result.
2000
2001 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
2002
2003 if (separate_io_proj) {
2004 // The caller requested separate projections be used by the fall
2005 // through and exceptional paths, so replace the projections for
2006 // the fall through path.
2007 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
2008 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
2009 }
2010
2011 // Capture the return value, if any.
2012 Node* ret;
2013 if (call->method() == nullptr || call->method()->return_type()->basic_type() == T_VOID) {
2014 ret = top();
2015 } else if (call->tf()->returns_inline_type_as_fields()) {
2016 // Return of multiple values (inline type fields): we create a
2017 // InlineType node, each field is a projection from the call.
2018 ciInlineKlass* vk = call->method()->return_type()->as_inline_klass();
2019 uint base_input = TypeFunc::Parms;
2020 ret = InlineTypeNode::make_from_multi(this, call, vk, base_input, false, false);
2021 } else {
2022 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
2023 ciType* t = call->method()->return_type();
2024 if (!t->is_loaded() && InlineTypeReturnedAsFields) {
2025 // The return type is unloaded but the callee might later be C2 compiled and then return
2026 // in scalarized form when the return type is loaded. Handle this similar to what we do in
2027 // PhaseMacroExpand::expand_mh_intrinsic_return by calling into the runtime to buffer.
2028 // It's a bit unfortunate because we will deopt anyway but the interpreter needs an oop.
2029 IdealKit ideal(this);
2030 IdealVariable res(ideal);
2031 ideal.declarations_done();
2032 ideal.if_then(ret, BoolTest::eq, ideal.makecon(TypePtr::NULL_PTR)); {
2033 // Return value is null
2034 ideal.set(res, ret);
2035 } ideal.else_(); {
2036 // Return value is non-null
2037 sync_kit(ideal);
2038
2039 // Change return type of call to scalarized return
2040 const TypeFunc* tf = call->_tf;
2041 const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc();
2042 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain);
2043 call->_tf = new_tf;
2044 _gvn.set_type(call, call->Value(&_gvn));
2045 _gvn.set_type(ret, ret->Value(&_gvn));
2046
2047 Node* store_to_buf_call = make_runtime_call(RC_NO_LEAF | RC_NO_IO,
2048 OptoRuntime::store_inline_type_fields_Type(),
2049 StubRoutines::store_inline_type_fields_to_buf(),
2050 nullptr, TypePtr::BOTTOM, ret);
2051
2052 // We don't know how many values are returned. This assumes the
2053 // worst case, that all available registers are used.
2054 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2055 if (domain->field_at(i) == Type::HALF) {
2056 store_to_buf_call->init_req(i, top());
2057 continue;
2058 }
2059 Node* proj =_gvn.transform(new ProjNode(call, i));
2060 store_to_buf_call->init_req(i, proj);
2061 }
2062 make_slow_call_ex(store_to_buf_call, env()->Throwable_klass(), false);
2063
2064 Node* buf = _gvn.transform(new ProjNode(store_to_buf_call, TypeFunc::Parms));
2065 const Type* buf_type = TypeOopPtr::make_from_klass(t->as_klass())->join_speculative(TypePtr::NOTNULL);
2066 buf = _gvn.transform(new CheckCastPPNode(control(), buf, buf_type));
2067
2068 ideal.set(res, buf);
2069 ideal.sync_kit(this);
2070 } ideal.end_if();
2071 sync_kit(ideal);
2072 ret = _gvn.transform(ideal.value(res));
2073 }
2074 if (t->is_klass()) {
2075 const Type* type = TypeOopPtr::make_from_klass(t->as_klass());
2076 if (type->is_inlinetypeptr()) {
2077 ret = InlineTypeNode::make_from_oop(this, ret, type->inline_klass());
2078 }
2079 }
2080 }
2081
2082 return ret;
2083 }
2084
2085 //--------------------set_predefined_input_for_runtime_call--------------------
2086 // Reading and setting the memory state is way conservative here.
2087 // The real problem is that I am not doing real Type analysis on memory,
2088 // so I cannot distinguish card mark stores from other stores. Across a GC
2089 // point the Store Barrier and the card mark memory has to agree. I cannot
2090 // have a card mark store and its barrier split across the GC point from
2091 // either above or below. Here I get that to happen by reading ALL of memory.
2092 // A better answer would be to separate out card marks from other memory.
2093 // For now, return the input memory state, so that it can be reused
2094 // after the call, if this call has restricted memory effects.
2095 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
2096 // Set fixed predefined input arguments
2097 Node* memory = reset_memory();
2098 Node* m = narrow_mem == nullptr ? memory : narrow_mem;
2099 call->init_req( TypeFunc::Control, control() );
2100 call->init_req( TypeFunc::I_O, top() ); // does no i/o
2101 call->init_req( TypeFunc::Memory, m ); // may gc ptrs
2152 if (use->is_MergeMem()) {
2153 wl.push(use);
2154 }
2155 }
2156 }
2157
2158 // Replace the call with the current state of the kit.
2159 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes, bool do_asserts) {
2160 JVMState* ejvms = nullptr;
2161 if (has_exceptions()) {
2162 ejvms = transfer_exceptions_into_jvms();
2163 }
2164
2165 ReplacedNodes replaced_nodes = map()->replaced_nodes();
2166 ReplacedNodes replaced_nodes_exception;
2167 Node* ex_ctl = top();
2168
2169 SafePointNode* final_state = stop();
2170
2171 // Find all the needed outputs of this call
2172 CallProjections* callprojs = call->extract_projections(true, do_asserts);
2173
2174 Unique_Node_List wl;
2175 Node* init_mem = call->in(TypeFunc::Memory);
2176 Node* final_mem = final_state->in(TypeFunc::Memory);
2177 Node* final_ctl = final_state->in(TypeFunc::Control);
2178 Node* final_io = final_state->in(TypeFunc::I_O);
2179
2180 // Replace all the old call edges with the edges from the inlining result
2181 if (callprojs->fallthrough_catchproj != nullptr) {
2182 C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl);
2183 }
2184 if (callprojs->fallthrough_memproj != nullptr) {
2185 if (final_mem->is_MergeMem()) {
2186 // Parser's exits MergeMem was not transformed but may be optimized
2187 final_mem = _gvn.transform(final_mem);
2188 }
2189 C->gvn_replace_by(callprojs->fallthrough_memproj, final_mem);
2190 add_mergemem_users_to_worklist(wl, final_mem);
2191 }
2192 if (callprojs->fallthrough_ioproj != nullptr) {
2193 C->gvn_replace_by(callprojs->fallthrough_ioproj, final_io);
2194 }
2195
2196 // Replace the result with the new result if it exists and is used
2197 if (callprojs->resproj[0] != nullptr && result != nullptr) {
2198 // If the inlined code is dead, the result projections for an inline type returned as
2199 // fields have not been replaced. They will go away once the call is replaced by TOP below.
2200 assert(callprojs->nb_resproj == 1 || (call->tf()->returns_inline_type_as_fields() && stopped()),
2201 "unexpected number of results");
2202 C->gvn_replace_by(callprojs->resproj[0], result);
2203 }
2204
2205 if (ejvms == nullptr) {
2206 // No exception edges to simply kill off those paths
2207 if (callprojs->catchall_catchproj != nullptr) {
2208 C->gvn_replace_by(callprojs->catchall_catchproj, C->top());
2209 }
2210 if (callprojs->catchall_memproj != nullptr) {
2211 C->gvn_replace_by(callprojs->catchall_memproj, C->top());
2212 }
2213 if (callprojs->catchall_ioproj != nullptr) {
2214 C->gvn_replace_by(callprojs->catchall_ioproj, C->top());
2215 }
2216 // Replace the old exception object with top
2217 if (callprojs->exobj != nullptr) {
2218 C->gvn_replace_by(callprojs->exobj, C->top());
2219 }
2220 } else {
2221 GraphKit ekit(ejvms);
2222
2223 // Load my combined exception state into the kit, with all phis transformed:
2224 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
2225 replaced_nodes_exception = ex_map->replaced_nodes();
2226
2227 Node* ex_oop = ekit.use_exception_state(ex_map);
2228
2229 if (callprojs->catchall_catchproj != nullptr) {
2230 C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control());
2231 ex_ctl = ekit.control();
2232 }
2233 if (callprojs->catchall_memproj != nullptr) {
2234 Node* ex_mem = ekit.reset_memory();
2235 C->gvn_replace_by(callprojs->catchall_memproj, ex_mem);
2236 add_mergemem_users_to_worklist(wl, ex_mem);
2237 }
2238 if (callprojs->catchall_ioproj != nullptr) {
2239 C->gvn_replace_by(callprojs->catchall_ioproj, ekit.i_o());
2240 }
2241
2242 // Replace the old exception object with the newly created one
2243 if (callprojs->exobj != nullptr) {
2244 C->gvn_replace_by(callprojs->exobj, ex_oop);
2245 }
2246 }
2247
2248 // Disconnect the call from the graph
2249 call->disconnect_inputs(C);
2250 C->gvn_replace_by(call, C->top());
2251
2252 // Clean up any MergeMems that feed other MergeMems since the
2253 // optimizer doesn't like that.
2254 while (wl.size() > 0) {
2255 _gvn.transform(wl.pop());
2256 }
2257
2258 if (callprojs->fallthrough_catchproj != nullptr && !final_ctl->is_top() && do_replaced_nodes) {
2259 replaced_nodes.apply(C, final_ctl);
2260 }
2261 if (!ex_ctl->is_top() && do_replaced_nodes) {
2262 replaced_nodes_exception.apply(C, ex_ctl);
2263 }
2264 }
2265
2266
2267 //------------------------------increment_counter------------------------------
2268 // for statistics: increment a VM counter by 1
2269
2270 void GraphKit::increment_counter(address counter_addr) {
2271 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2272 increment_counter(adr1);
2273 }
2274
2275 void GraphKit::increment_counter(Node* counter_addr) {
2276 Node* ctrl = control();
2277 Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, MemNode::unordered);
2278 Node* incr = _gvn.transform(new AddLNode(cnt, _gvn.longcon(1)));
2438 *
2439 * @param n node that the type applies to
2440 * @param exact_kls type from profiling
2441 * @param maybe_null did profiling see null?
2442 *
2443 * @return node with improved type
2444 */
2445 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2446 const Type* current_type = _gvn.type(n);
2447 assert(UseTypeSpeculation, "type speculation must be on");
2448
2449 const TypePtr* speculative = current_type->speculative();
2450
2451 // Should the klass from the profile be recorded in the speculative type?
2452 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2453 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls, Type::trust_interfaces);
2454 const TypeOopPtr* xtype = tklass->as_instance_type();
2455 assert(xtype->klass_is_exact(), "Should be exact");
2456 // Any reason to believe n is not null (from this profiling or a previous one)?
2457 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2458 const TypePtr* ptr = (ptr_kind != ProfileNeverNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2459 // record the new speculative type's depth
2460 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2461 speculative = speculative->with_inline_depth(jvms()->depth());
2462 } else if (current_type->would_improve_ptr(ptr_kind)) {
2463 // Profiling report that null was never seen so we can change the
2464 // speculative type to non null ptr.
2465 if (ptr_kind == ProfileAlwaysNull) {
2466 speculative = TypePtr::NULL_PTR;
2467 } else {
2468 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2469 const TypePtr* ptr = TypePtr::NOTNULL;
2470 if (speculative != nullptr) {
2471 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2472 } else {
2473 speculative = ptr;
2474 }
2475 }
2476 }
2477
2478 if (speculative != current_type->speculative()) {
2479 // Build a type with a speculative type (what we think we know
2480 // about the type but will need a guard when we use it)
2481 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative);
2482 // We're changing the type, we need a new CheckCast node to carry
2483 // the new type. The new type depends on the control: what
2484 // profiling tells us is only valid from here as far as we can
2485 // tell.
2486 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2487 cast = _gvn.transform(cast);
2488 replace_in_map(n, cast);
2489 n = cast;
2490 }
2491
2492 return n;
2493 }
2494
2495 /**
2496 * Record profiling data from receiver profiling at an invoke with the
2497 * type system so that it can propagate it (speculation)
2498 *
2499 * @param n receiver node
2500 *
2501 * @return node with improved type
2502 */
2503 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2504 if (!UseTypeSpeculation) {
2505 return n;
2506 }
2507 ciKlass* exact_kls = profile_has_unique_klass();
2508 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2509 if ((java_bc() == Bytecodes::_checkcast ||
2510 java_bc() == Bytecodes::_instanceof ||
2511 java_bc() == Bytecodes::_aastore) &&
2512 method()->method_data()->is_mature()) {
2513 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2514 if (data != nullptr) {
2515 if (java_bc() == Bytecodes::_aastore) {
2516 ciKlass* array_type = nullptr;
2517 ciKlass* element_type = nullptr;
2518 ProfilePtrKind element_ptr = ProfileMaybeNull;
2519 bool flat_array = true;
2520 bool null_free_array = true;
2521 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
2522 exact_kls = element_type;
2523 ptr_kind = element_ptr;
2524 } else {
2525 if (!data->as_BitData()->null_seen()) {
2526 ptr_kind = ProfileNeverNull;
2527 } else {
2528 assert(data->is_ReceiverTypeData(), "bad profile data type");
2529 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2530 uint i = 0;
2531 for (; i < call->row_limit(); i++) {
2532 ciKlass* receiver = call->receiver(i);
2533 if (receiver != nullptr) {
2534 break;
2535 }
2536 }
2537 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2538 }
2539 }
2540 }
2541 }
2542 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2543 }
2544
2545 /**
2546 * Record profiling data from argument profiling at an invoke with the
2547 * type system so that it can propagate it (speculation)
2548 *
2549 * @param dest_method target method for the call
2550 * @param bc what invoke bytecode is this?
2551 */
2552 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2553 if (!UseTypeSpeculation) {
2554 return;
2555 }
2556 const TypeFunc* tf = TypeFunc::make(dest_method);
2557 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2558 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2559 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2560 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2561 if (is_reference_type(targ->basic_type())) {
2562 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2563 ciKlass* better_type = nullptr;
2564 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2565 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2566 }
2567 i++;
2568 }
2569 }
2570 }
2571
2572 /**
2573 * Record profiling data from parameter profiling at an invoke with
2574 * the type system so that it can propagate it (speculation)
2575 */
2576 void GraphKit::record_profiled_parameters_for_speculation() {
2577 if (!UseTypeSpeculation) {
2578 return;
2579 }
2580 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2700 // The first null ends the list.
2701 Node* parm0, Node* parm1,
2702 Node* parm2, Node* parm3,
2703 Node* parm4, Node* parm5,
2704 Node* parm6, Node* parm7) {
2705 assert(call_addr != nullptr, "must not call null targets");
2706
2707 // Slow-path call
2708 bool is_leaf = !(flags & RC_NO_LEAF);
2709 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2710 if (call_name == nullptr) {
2711 assert(!is_leaf, "must supply name for leaf");
2712 call_name = OptoRuntime::stub_name(call_addr);
2713 }
2714 CallNode* call;
2715 if (!is_leaf) {
2716 call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type);
2717 } else if (flags & RC_NO_FP) {
2718 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2719 } else if (flags & RC_VECTOR){
2720 uint num_bits = call_type->range_sig()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte;
2721 call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits);
2722 } else {
2723 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2724 }
2725
2726 // The following is similar to set_edges_for_java_call,
2727 // except that the memory effects of the call are restricted to AliasIdxRaw.
2728
2729 // Slow path call has no side-effects, uses few values
2730 bool wide_in = !(flags & RC_NARROW_MEM);
2731 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2732
2733 Node* prev_mem = nullptr;
2734 if (wide_in) {
2735 prev_mem = set_predefined_input_for_runtime_call(call);
2736 } else {
2737 assert(!wide_out, "narrow in => narrow out");
2738 Node* narrow_mem = memory(adr_type);
2739 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2740 }
2741
2742 // Hook each parm in order. Stop looking at the first null.
2743 if (parm0 != nullptr) { call->init_req(TypeFunc::Parms+0, parm0);
2744 if (parm1 != nullptr) { call->init_req(TypeFunc::Parms+1, parm1);
2745 if (parm2 != nullptr) { call->init_req(TypeFunc::Parms+2, parm2);
2746 if (parm3 != nullptr) { call->init_req(TypeFunc::Parms+3, parm3);
2747 if (parm4 != nullptr) { call->init_req(TypeFunc::Parms+4, parm4);
2748 if (parm5 != nullptr) { call->init_req(TypeFunc::Parms+5, parm5);
2749 if (parm6 != nullptr) { call->init_req(TypeFunc::Parms+6, parm6);
2750 if (parm7 != nullptr) { call->init_req(TypeFunc::Parms+7, parm7);
2751 /* close each nested if ===> */ } } } } } } } }
2752 assert(call->in(call->req()-1) != nullptr || (call->req()-1) > (TypeFunc::Parms+7), "must initialize all parms");
2753
2754 if (!is_leaf) {
2755 // Non-leaves can block and take safepoints:
2756 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2757 }
2758 // Non-leaves can throw exceptions:
2759 if (has_io) {
2760 call->set_req(TypeFunc::I_O, i_o());
2761 }
2762
2763 if (flags & RC_UNCOMMON) {
2764 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2765 // (An "if" probability corresponds roughly to an unconditional count.
2766 // Sort of.)
2767 call->set_cnt(PROB_UNLIKELY_MAG(4));
2768 }
2769
2770 Node* c = _gvn.transform(call);
2771 assert(c == call, "cannot disappear");
2772
2780
2781 if (has_io) {
2782 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2783 }
2784 return call;
2785
2786 }
2787
2788 // i2b
2789 Node* GraphKit::sign_extend_byte(Node* in) {
2790 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24)));
2791 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24)));
2792 }
2793
2794 // i2s
2795 Node* GraphKit::sign_extend_short(Node* in) {
2796 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16)));
2797 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16)));
2798 }
2799
2800
2801 //------------------------------merge_memory-----------------------------------
2802 // Merge memory from one path into the current memory state.
2803 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2804 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2805 Node* old_slice = mms.force_memory();
2806 Node* new_slice = mms.memory2();
2807 if (old_slice != new_slice) {
2808 PhiNode* phi;
2809 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2810 if (mms.is_empty()) {
2811 // clone base memory Phi's inputs for this memory slice
2812 assert(old_slice == mms.base_memory(), "sanity");
2813 phi = PhiNode::make(region, nullptr, Type::MEMORY, mms.adr_type(C));
2814 _gvn.set_type(phi, Type::MEMORY);
2815 for (uint i = 1; i < phi->req(); i++) {
2816 phi->init_req(i, old_slice->in(i));
2817 }
2818 } else {
2819 phi = old_slice->as_Phi(); // Phi was generated already
2820 }
2877 gvn.transform(iff);
2878 if (!bol->is_Con()) gvn.record_for_igvn(iff);
2879 return iff;
2880 }
2881
2882 //-------------------------------gen_subtype_check-----------------------------
2883 // Generate a subtyping check. Takes as input the subtype and supertype.
2884 // Returns 2 values: sets the default control() to the true path and returns
2885 // the false path. Only reads invariant memory; sets no (visible) memory.
2886 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2887 // but that's not exposed to the optimizer. This call also doesn't take in an
2888 // Object; if you wish to check an Object you need to load the Object's class
2889 // prior to coming here.
2890 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn,
2891 ciMethod* method, int bci) {
2892 Compile* C = gvn.C;
2893 if ((*ctrl)->is_top()) {
2894 return C->top();
2895 }
2896
2897 const TypeKlassPtr* klass_ptr_type = gvn.type(superklass)->is_klassptr();
2898 const TypeAryKlassPtr* ary_klass_t = klass_ptr_type->isa_aryklassptr();
2899 Node* vm_superklass = superklass;
2900 // TODO 8366668 Compute the VM type here for when we do a direct pointer comparison
2901 if (ary_klass_t && ary_klass_t->klass_is_exact() && ary_klass_t->exact_klass()->is_obj_array_klass()) {
2902 ary_klass_t = ary_klass_t->get_vm_type();
2903 vm_superklass = gvn.makecon(ary_klass_t);
2904 }
2905
2906 // Fast check for identical types, perhaps identical constants.
2907 // The types can even be identical non-constants, in cases
2908 // involving Array.newInstance, Object.clone, etc.
2909 if (subklass == superklass)
2910 return C->top(); // false path is dead; no test needed.
2911
2912 if (gvn.type(superklass)->singleton()) {
2913 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
2914 const TypeKlassPtr* subk = gvn.type(subklass)->is_klassptr();
2915
2916 // In the common case of an exact superklass, try to fold up the
2917 // test before generating code. You may ask, why not just generate
2918 // the code and then let it fold up? The answer is that the generated
2919 // code will necessarily include null checks, which do not always
2920 // completely fold away. If they are also needless, then they turn
2921 // into a performance loss. Example:
2922 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2923 // Here, the type of 'fa' is often exact, so the store check
2924 // of fa[1]=x will fold up, without testing the nullness of x.
2925 //
2926 // At macro expansion, we would have already folded the SubTypeCheckNode
2927 // being expanded here because we always perform the static sub type
2928 // check in SubTypeCheckNode::sub() regardless of whether
2929 // StressReflectiveCode is set or not. We can therefore skip this
2930 // static check when StressReflectiveCode is on.
2931 switch (C->static_subtype_check(superk, subk)) {
2932 case Compile::SSC_always_false:
2933 {
2934 Node* always_fail = *ctrl;
2935 *ctrl = gvn.C->top();
2936 return always_fail;
2937 }
2938 case Compile::SSC_always_true:
2939 return C->top();
2940 case Compile::SSC_easy_test:
2941 {
2942 // Just do a direct pointer compare and be done.
2943 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, vm_superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2944 *ctrl = gvn.transform(new IfTrueNode(iff));
2945 return gvn.transform(new IfFalseNode(iff));
2946 }
2947 case Compile::SSC_full_test:
2948 break;
2949 default:
2950 ShouldNotReachHere();
2951 }
2952 }
2953
2954 // %%% Possible further optimization: Even if the superklass is not exact,
2955 // if the subklass is the unique subtype of the superklass, the check
2956 // will always succeed. We could leave a dependency behind to ensure this.
2957
2958 // First load the super-klass's check-offset
2959 Node *p1 = gvn.transform(new AddPNode(superklass, superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2960 Node* m = C->immutable_memory();
2961 Node *chk_off = gvn.transform(new LoadINode(nullptr, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2962 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2963 const TypeInt* chk_off_t = chk_off->Value(&gvn)->isa_int();
2964 int chk_off_con = (chk_off_t != nullptr && chk_off_t->is_con()) ? chk_off_t->get_con() : cacheoff_con;
2965 // TODO 8366668 Re-enable. This breaks test/hotspot/jtreg/compiler/c2/irTests/ProfileAtTypeCheck.java
2966 bool might_be_cache = true;//(chk_off_con == cacheoff_con);
2967
2968 // Load from the sub-klass's super-class display list, or a 1-word cache of
2969 // the secondary superclass list, or a failing value with a sentinel offset
2970 // if the super-klass is an interface or exceptionally deep in the Java
2971 // hierarchy and we have to scan the secondary superclass list the hard way.
2972 // Worst-case type is a little odd: null is allowed as a result (usually
2973 // klass loads can never produce a null).
2974 Node *chk_off_X = chk_off;
2975 #ifdef _LP64
2976 chk_off_X = gvn.transform(new ConvI2LNode(chk_off_X));
2977 #endif
2978 Node *p2 = gvn.transform(new AddPNode(subklass,subklass,chk_off_X));
2979 // For some types like interfaces the following loadKlass is from a 1-word
2980 // cache which is mutable so can't use immutable memory. Other
2981 // types load from the super-class display table which is immutable.
2982 Node *kmem = C->immutable_memory();
2983 // secondary_super_cache is not immutable but can be treated as such because:
2984 // - no ideal node writes to it in a way that could cause an
2985 // incorrect/missed optimization of the following Load.
2986 // - it's a cache so, worse case, not reading the latest value
2997
2998 // Gather the various success & failures here
2999 RegionNode* r_not_subtype = new RegionNode(3);
3000 gvn.record_for_igvn(r_not_subtype);
3001 RegionNode* r_ok_subtype = new RegionNode(4);
3002 gvn.record_for_igvn(r_ok_subtype);
3003
3004 // If we might perform an expensive check, first try to take advantage of profile data that was attached to the
3005 // SubTypeCheck node
3006 if (might_be_cache && method != nullptr && VM_Version::profile_all_receivers_at_type_check()) {
3007 ciCallProfile profile = method->call_profile_at_bci(bci);
3008 float total_prob = 0;
3009 for (int i = 0; profile.has_receiver(i); ++i) {
3010 float prob = profile.receiver_prob(i);
3011 total_prob += prob;
3012 }
3013 if (total_prob * 100. >= TypeProfileSubTypeCheckCommonThreshold) {
3014 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
3015 for (int i = 0; profile.has_receiver(i); ++i) {
3016 ciKlass* klass = profile.receiver(i);
3017 // TODO 8366668 Do we need adjustments here??
3018 const TypeKlassPtr* klass_t = TypeKlassPtr::make(klass);
3019 Compile::SubTypeCheckResult result = C->static_subtype_check(superk, klass_t);
3020 if (result != Compile::SSC_always_true && result != Compile::SSC_always_false) {
3021 continue;
3022 }
3023 float prob = profile.receiver_prob(i);
3024 ConNode* klass_node = gvn.makecon(klass_t);
3025 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, klass_node, BoolTest::eq, prob, gvn, T_ADDRESS);
3026 Node* iftrue = gvn.transform(new IfTrueNode(iff));
3027
3028 if (result == Compile::SSC_always_true) {
3029 r_ok_subtype->add_req(iftrue);
3030 } else {
3031 assert(result == Compile::SSC_always_false, "");
3032 r_not_subtype->add_req(iftrue);
3033 }
3034 *ctrl = gvn.transform(new IfFalseNode(iff));
3035 }
3036 }
3037 }
3048 // subklass. In this case we need exactly the 1 test above and we can
3049 // return those results immediately.
3050 if (!might_be_cache) {
3051 Node* not_subtype_ctrl = *ctrl;
3052 *ctrl = iftrue1; // We need exactly the 1 test above
3053 PhaseIterGVN* igvn = gvn.is_IterGVN();
3054 if (igvn != nullptr) {
3055 igvn->remove_globally_dead_node(r_ok_subtype);
3056 igvn->remove_globally_dead_node(r_not_subtype);
3057 }
3058 return not_subtype_ctrl;
3059 }
3060
3061 r_ok_subtype->init_req(1, iftrue1);
3062
3063 // Check for immediate negative hit. Happens roughly 11% of the time (which
3064 // is roughly 63% of the remaining cases). Test to see if the loaded
3065 // check-offset points into the subklass display list or the 1-element
3066 // cache. If it points to the display (and NOT the cache) and the display
3067 // missed then it's not a subtype.
3068 // TODO 8366668 Re-enable
3069 /*
3070 Node *cacheoff = gvn.intcon(cacheoff_con);
3071 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
3072 r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2)));
3073 *ctrl = gvn.transform(new IfFalseNode(iff2));
3074 */
3075 // Check for self. Very rare to get here, but it is taken 1/3 the time.
3076 // No performance impact (too rare) but allows sharing of secondary arrays
3077 // which has some footprint reduction.
3078 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, vm_superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
3079 r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3)));
3080 *ctrl = gvn.transform(new IfFalseNode(iff3));
3081
3082 // -- Roads not taken here: --
3083 // We could also have chosen to perform the self-check at the beginning
3084 // of this code sequence, as the assembler does. This would not pay off
3085 // the same way, since the optimizer, unlike the assembler, can perform
3086 // static type analysis to fold away many successful self-checks.
3087 // Non-foldable self checks work better here in second position, because
3088 // the initial primary superclass check subsumes a self-check for most
3089 // types. An exception would be a secondary type like array-of-interface,
3090 // which does not appear in its own primary supertype display.
3091 // Finally, we could have chosen to move the self-check into the
3092 // PartialSubtypeCheckNode, and from there out-of-line in a platform
3093 // dependent manner. But it is worthwhile to have the check here,
3094 // where it can be perhaps be optimized. The cost in code space is
3095 // small (register compare, branch).
3096
3097 // Now do a linear scan of the secondary super-klass array. Again, no real
3098 // performance impact (too rare) but it's gotta be done.
3099 // Since the code is rarely used, there is no penalty for moving it
3100 // out of line, and it can only improve I-cache density.
3101 // The decision to inline or out-of-line this final check is platform
3102 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
3103 Node* psc = gvn.transform(
3104 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
3105
3106 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
3107 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
3108 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
3109
3110 // Return false path; set default control to true path.
3111 *ctrl = gvn.transform(r_ok_subtype);
3112 return gvn.transform(r_not_subtype);
3113 }
3114
3115 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
3116 const Type* sub_t = _gvn.type(obj_or_subklass);
3117 if (sub_t->make_oopptr() != nullptr && sub_t->make_oopptr()->is_inlinetypeptr()) {
3118 sub_t = TypeKlassPtr::make(sub_t->inline_klass());
3119 obj_or_subklass = makecon(sub_t);
3120 }
3121 bool expand_subtype_check = C->post_loop_opts_phase(); // macro node expansion is over
3122 if (expand_subtype_check) {
3123 MergeMemNode* mem = merged_memory();
3124 Node* ctrl = control();
3125 Node* subklass = obj_or_subklass;
3126 if (!sub_t->isa_klassptr()) {
3127 subklass = load_object_klass(obj_or_subklass);
3128 }
3129
3130 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn, method(), bci());
3131 set_control(ctrl);
3132 return n;
3133 }
3134
3135 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass, method(), bci()));
3136 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
3137 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3138 set_control(_gvn.transform(new IfTrueNode(iff)));
3139 return _gvn.transform(new IfFalseNode(iff));
3140 }
3141
3142 // Profile-driven exact type check:
3143 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
3144 float prob, Node* *casted_receiver) {
3145 assert(!klass->is_interface(), "no exact type check on interfaces");
3146 Node* fail = top();
3147 const Type* rec_t = _gvn.type(receiver);
3148 if (rec_t->is_inlinetypeptr()) {
3149 if (klass->equals(rec_t->inline_klass())) {
3150 (*casted_receiver) = receiver; // Always passes
3151 } else {
3152 (*casted_receiver) = top(); // Always fails
3153 fail = control();
3154 set_control(top());
3155 }
3156 return fail;
3157 }
3158 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces);
3159 const TypeAryKlassPtr* ary_klass_t = tklass->isa_aryklassptr();
3160 // TODO 8366668 Compute the VM type
3161 if (ary_klass_t && ary_klass_t->klass_is_exact() && ary_klass_t->exact_klass()->is_obj_array_klass()) {
3162 tklass = ary_klass_t->get_vm_type();
3163 }
3164 Node* recv_klass = load_object_klass(receiver);
3165 fail = type_check(recv_klass, tklass, prob);
3166
3167 if (!stopped()) {
3168 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3169 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
3170 assert(recv_xtype->klass_is_exact(), "");
3171
3172 if (!receiver_type->higher_equal(recv_xtype)) { // ignore redundant casts
3173 // Subsume downstream occurrences of receiver with a cast to
3174 // recv_xtype, since now we know what the type will be.
3175 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
3176 Node* res = _gvn.transform(cast);
3177 if (recv_xtype->is_inlinetypeptr()) {
3178 assert(!gvn().type(res)->maybe_null(), "receiver should never be null");
3179 res = InlineTypeNode::make_from_oop(this, res, recv_xtype->inline_klass());
3180 }
3181 (*casted_receiver) = res;
3182 assert(!(*casted_receiver)->is_top(), "that path should be unreachable");
3183 // (User must make the replace_in_map call.)
3184 }
3185 }
3186
3187 return fail;
3188 }
3189
3190 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass,
3191 float prob) {
3192 Node* want_klass = makecon(tklass);
3193 Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass));
3194 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3195 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
3196 set_control(_gvn.transform(new IfTrueNode (iff)));
3197 Node* fail = _gvn.transform(new IfFalseNode(iff));
3198 return fail;
3199 }
3200
3201 //------------------------------subtype_check_receiver-------------------------
3202 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
3203 Node** casted_receiver) {
3204 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces)->try_improve();
3205 Node* want_klass = makecon(tklass);
3206
3207 Node* slow_ctl = gen_subtype_check(receiver, want_klass);
3208
3209 // Ignore interface type information until interface types are properly tracked.
3210 if (!stopped() && !klass->is_interface()) {
3211 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3212 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
3213 if (receiver_type != nullptr && !receiver_type->higher_equal(recv_type)) { // ignore redundant casts
3214 Node* cast = _gvn.transform(new CheckCastPPNode(control(), receiver, recv_type));
3215 if (recv_type->is_inlinetypeptr()) {
3216 cast = InlineTypeNode::make_from_oop(this, cast, recv_type->inline_klass());
3217 }
3218 (*casted_receiver) = cast;
3219 }
3220 }
3221
3222 return slow_ctl;
3223 }
3224
3225 //------------------------------seems_never_null-------------------------------
3226 // Use null_seen information if it is available from the profile.
3227 // If we see an unexpected null at a type check we record it and force a
3228 // recompile; the offending check will be recompiled to handle nulls.
3229 // If we see several offending BCIs, then all checks in the
3230 // method will be recompiled.
3231 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
3232 speculating = !_gvn.type(obj)->speculative_maybe_null();
3233 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
3234 if (UncommonNullCast // Cutout for this technique
3235 && obj != null() // And not the -Xcomp stupid case?
3236 && !too_many_traps(reason)
3237 ) {
3238 if (speculating) {
3307
3308 //------------------------maybe_cast_profiled_receiver-------------------------
3309 // If the profile has seen exactly one type, narrow to exactly that type.
3310 // Subsequent type checks will always fold up.
3311 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3312 const TypeKlassPtr* require_klass,
3313 ciKlass* spec_klass,
3314 bool safe_for_replace) {
3315 if (!UseTypeProfile || !TypeProfileCasts) return nullptr;
3316
3317 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != nullptr);
3318
3319 // Make sure we haven't already deoptimized from this tactic.
3320 if (too_many_traps_or_recompiles(reason))
3321 return nullptr;
3322
3323 // (No, this isn't a call, but it's enough like a virtual call
3324 // to use the same ciMethod accessor to get the profile info...)
3325 // If we have a speculative type use it instead of profiling (which
3326 // may not help us)
3327 ciKlass* exact_kls = spec_klass;
3328 if (exact_kls == nullptr) {
3329 if (java_bc() == Bytecodes::_aastore) {
3330 ciKlass* array_type = nullptr;
3331 ciKlass* element_type = nullptr;
3332 ProfilePtrKind element_ptr = ProfileMaybeNull;
3333 bool flat_array = true;
3334 bool null_free_array = true;
3335 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
3336 exact_kls = element_type;
3337 } else {
3338 exact_kls = profile_has_unique_klass();
3339 }
3340 }
3341 if (exact_kls != nullptr) {// no cast failures here
3342 if (require_klass == nullptr ||
3343 C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls, Type::trust_interfaces)) == Compile::SSC_always_true) {
3344 // If we narrow the type to match what the type profile sees or
3345 // the speculative type, we can then remove the rest of the
3346 // cast.
3347 // This is a win, even if the exact_kls is very specific,
3348 // because downstream operations, such as method calls,
3349 // will often benefit from the sharper type.
3350 Node* exact_obj = not_null_obj; // will get updated in place...
3351 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3352 &exact_obj);
3353 { PreserveJVMState pjvms(this);
3354 set_control(slow_ctl);
3355 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3356 }
3357 if (safe_for_replace) {
3358 replace_in_map(not_null_obj, exact_obj);
3359 }
3360 return exact_obj;
3450 // If not_null_obj is dead, only null-path is taken
3451 if (stopped()) { // Doing instance-of on a null?
3452 set_control(null_ctl);
3453 return intcon(0);
3454 }
3455 region->init_req(_null_path, null_ctl);
3456 phi ->init_req(_null_path, intcon(0)); // Set null path value
3457 if (null_ctl == top()) {
3458 // Do this eagerly, so that pattern matches like is_diamond_phi
3459 // will work even during parsing.
3460 assert(_null_path == PATH_LIMIT-1, "delete last");
3461 region->del_req(_null_path);
3462 phi ->del_req(_null_path);
3463 }
3464
3465 // Do we know the type check always succeed?
3466 bool known_statically = false;
3467 if (_gvn.type(superklass)->singleton()) {
3468 const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr();
3469 const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type();
3470 if (subk != nullptr && subk->is_loaded()) {
3471 int static_res = C->static_subtype_check(superk, subk);
3472 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3473 }
3474 }
3475
3476 if (!known_statically) {
3477 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3478 // We may not have profiling here or it may not help us. If we
3479 // have a speculative type use it to perform an exact cast.
3480 ciKlass* spec_obj_type = obj_type->speculative_type();
3481 if (spec_obj_type != nullptr || (ProfileDynamicTypes && data != nullptr)) {
3482 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, nullptr, spec_obj_type, safe_for_replace);
3483 if (stopped()) { // Profile disagrees with this path.
3484 set_control(null_ctl); // Null is the only remaining possibility.
3485 return intcon(0);
3486 }
3487 if (cast_obj != nullptr) {
3488 not_null_obj = cast_obj;
3489 }
3490 }
3506 record_for_igvn(region);
3507
3508 // If we know the type check always succeeds then we don't use the
3509 // profiling data at this bytecode. Don't lose it, feed it to the
3510 // type system as a speculative type.
3511 if (safe_for_replace) {
3512 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3513 replace_in_map(obj, casted_obj);
3514 }
3515
3516 return _gvn.transform(phi);
3517 }
3518
3519 //-------------------------------gen_checkcast---------------------------------
3520 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3521 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3522 // uncommon-trap paths work. Adjust stack after this call.
3523 // If failure_control is supplied and not null, it is filled in with
3524 // the control edge for the cast failure. Otherwise, an appropriate
3525 // uncommon trap or exception is thrown.
3526 Node* GraphKit::gen_checkcast(Node* obj, Node* superklass, Node* *failure_control, bool null_free, bool maybe_larval) {
3527 kill_dead_locals(); // Benefit all the uncommon traps
3528 const TypeKlassPtr* klass_ptr_type = _gvn.type(superklass)->is_klassptr();
3529 const Type* obj_type = _gvn.type(obj);
3530 if (obj_type->is_inlinetypeptr() && !obj_type->maybe_null() && klass_ptr_type->klass_is_exact() && obj_type->inline_klass() == klass_ptr_type->exact_klass(true)) {
3531 // Special case: larval inline objects must not be scalarized. They are also generally not
3532 // allowed to participate in most operations except as the first operand of putfield, or as an
3533 // argument to a constructor invocation with it being a receiver, Unsafe::putXXX with it being
3534 // the first argument, or Unsafe::finishPrivateBuffer. This allows us to aggressively scalarize
3535 // value objects in all other places. This special case comes from the limitation of the Java
3536 // language, Unsafe::makePrivateBuffer returns an Object that is checkcast-ed to the concrete
3537 // value type. We must do this first because C->static_subtype_check may do nothing when
3538 // StressReflectiveCode is set.
3539 return obj;
3540 }
3541
3542 // Else it must be a non-larval object
3543 obj = cast_to_non_larval(obj);
3544
3545 const TypeKlassPtr* improved_klass_ptr_type = klass_ptr_type->try_improve();
3546 const TypeOopPtr* toop = improved_klass_ptr_type->cast_to_exactness(false)->as_instance_type();
3547 bool safe_for_replace = (failure_control == nullptr);
3548 assert(!null_free || toop->can_be_inline_type(), "must be an inline type pointer");
3549
3550 // Fast cutout: Check the case that the cast is vacuously true.
3551 // This detects the common cases where the test will short-circuit
3552 // away completely. We do this before we perform the null check,
3553 // because if the test is going to turn into zero code, we don't
3554 // want a residual null check left around. (Causes a slowdown,
3555 // for example, in some objArray manipulations, such as a[i]=a[j].)
3556 if (improved_klass_ptr_type->singleton()) {
3557 const TypeKlassPtr* kptr = nullptr;
3558 if (obj_type->isa_oop_ptr()) {
3559 kptr = obj_type->is_oopptr()->as_klass_type();
3560 } else if (obj->is_InlineType()) {
3561 ciInlineKlass* vk = obj_type->inline_klass();
3562 kptr = TypeInstKlassPtr::make(TypePtr::NotNull, vk, Type::Offset(0));
3563 }
3564
3565 if (kptr != nullptr) {
3566 switch (C->static_subtype_check(improved_klass_ptr_type, kptr)) {
3567 case Compile::SSC_always_true:
3568 // If we know the type check always succeed then we don't use
3569 // the profiling data at this bytecode. Don't lose it, feed it
3570 // to the type system as a speculative type.
3571 obj = record_profiled_receiver_for_speculation(obj);
3572 if (null_free) {
3573 assert(safe_for_replace, "must be");
3574 obj = null_check(obj);
3575 }
3576 assert(stopped() || !toop->is_inlinetypeptr() || obj->is_InlineType(), "should have been scalarized");
3577 return obj;
3578 case Compile::SSC_always_false:
3579 if (null_free) {
3580 assert(safe_for_replace, "must be");
3581 obj = null_check(obj);
3582 }
3583 // It needs a null check because a null will *pass* the cast check.
3584 if (obj_type->isa_oopptr() != nullptr && !obj_type->is_oopptr()->maybe_null()) {
3585 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3586 Deoptimization::DeoptReason reason = is_aastore ?
3587 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3588 builtin_throw(reason);
3589 return top();
3590 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3591 return null_assert(obj);
3592 }
3593 break; // Fall through to full check
3594 default:
3595 break;
3596 }
3597 }
3598 }
3599
3600 ciProfileData* data = nullptr;
3601 if (failure_control == nullptr) { // use MDO in regular case only
3602 assert(java_bc() == Bytecodes::_aastore ||
3603 java_bc() == Bytecodes::_checkcast,
3604 "interpreter profiles type checks only for these BCs");
3605 if (method()->method_data()->is_mature()) {
3606 data = method()->method_data()->bci_to_data(bci());
3607 }
3608 }
3609
3610 // Make the merge point
3611 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3612 RegionNode* region = new RegionNode(PATH_LIMIT);
3613 Node* phi = new PhiNode(region, toop);
3614 _gvn.set_type(region, Type::CONTROL);
3615 _gvn.set_type(phi, toop);
3616
3617 C->set_has_split_ifs(true); // Has chance for split-if optimization
3618
3619 // Use null-cast information if it is available
3620 bool speculative_not_null = false;
3621 bool never_see_null = ((failure_control == nullptr) // regular case only
3622 && seems_never_null(obj, data, speculative_not_null));
3623
3624 if (obj->is_InlineType()) {
3625 // Re-execute if buffering during triggers deoptimization
3626 PreserveReexecuteState preexecs(this);
3627 jvms()->set_should_reexecute(true);
3628 obj = obj->as_InlineType()->buffer(this, safe_for_replace);
3629 }
3630
3631 // Null check; get casted pointer; set region slot 3
3632 Node* null_ctl = top();
3633 Node* not_null_obj = nullptr;
3634 if (null_free) {
3635 assert(safe_for_replace, "must be");
3636 not_null_obj = null_check(obj);
3637 } else {
3638 not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3639 }
3640
3641 // If not_null_obj is dead, only null-path is taken
3642 if (stopped()) { // Doing instance-of on a null?
3643 set_control(null_ctl);
3644 if (toop->is_inlinetypeptr()) {
3645 return InlineTypeNode::make_null(_gvn, toop->inline_klass());
3646 }
3647 return null();
3648 }
3649 region->init_req(_null_path, null_ctl);
3650 phi ->init_req(_null_path, null()); // Set null path value
3651 if (null_ctl == top()) {
3652 // Do this eagerly, so that pattern matches like is_diamond_phi
3653 // will work even during parsing.
3654 assert(_null_path == PATH_LIMIT-1, "delete last");
3655 region->del_req(_null_path);
3656 phi ->del_req(_null_path);
3657 }
3658
3659 Node* cast_obj = nullptr;
3660 if (improved_klass_ptr_type->klass_is_exact()) {
3661 // The following optimization tries to statically cast the speculative type of the object
3662 // (for example obtained during profiling) to the type of the superklass and then do a
3663 // dynamic check that the type of the object is what we expect. To work correctly
3664 // for checkcast and aastore the type of superklass should be exact.
3665 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3666 // We may not have profiling here or it may not help us. If we have
3667 // a speculative type use it to perform an exact cast.
3668 ciKlass* spec_obj_type = obj_type->speculative_type();
3669 if (spec_obj_type != nullptr || data != nullptr) {
3670 cast_obj = maybe_cast_profiled_receiver(not_null_obj, improved_klass_ptr_type, spec_obj_type, safe_for_replace);
3671 if (cast_obj != nullptr) {
3672 if (failure_control != nullptr) // failure is now impossible
3673 (*failure_control) = top();
3674 // adjust the type of the phi to the exact klass:
3675 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3676 }
3677 }
3678 }
3679
3680 if (cast_obj == nullptr) {
3681 // Generate the subtype check
3682 Node* improved_superklass = superklass;
3683 if (improved_klass_ptr_type != klass_ptr_type && improved_klass_ptr_type->singleton()) {
3684 // Only improve the super class for constants which allows subsequent sub type checks to possibly be commoned up.
3685 // The other non-constant cases cannot be improved with a cast node here since they could be folded to top.
3686 // Additionally, the benefit would only be minor in non-constant cases.
3687 improved_superklass = makecon(improved_klass_ptr_type);
3688 }
3689 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, improved_superklass);
3690 // Plug in success path into the merge
3691 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3692 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3693 if (failure_control == nullptr) {
3694 if (not_subtype_ctrl != top()) { // If failure is possible
3695 PreserveJVMState pjvms(this);
3696 set_control(not_subtype_ctrl);
3697 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3698 Deoptimization::DeoptReason reason = is_aastore ?
3699 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3700 builtin_throw(reason);
3701 }
3702 } else {
3703 (*failure_control) = not_subtype_ctrl;
3704 }
3705 }
3706
3707 region->init_req(_obj_path, control());
3708 phi ->init_req(_obj_path, cast_obj);
3709
3710 // A merge of null or Casted-NotNull obj
3711 Node* res = _gvn.transform(phi);
3712
3713 // Note I do NOT always 'replace_in_map(obj,result)' here.
3714 // if( tk->klass()->can_be_primary_super() )
3715 // This means that if I successfully store an Object into an array-of-String
3716 // I 'forget' that the Object is really now known to be a String. I have to
3717 // do this because we don't have true union types for interfaces - if I store
3718 // a Baz into an array-of-Interface and then tell the optimizer it's an
3719 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3720 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3721 // replace_in_map( obj, res );
3722
3723 // Return final merged results
3724 set_control( _gvn.transform(region) );
3725 record_for_igvn(region);
3726
3727 bool not_inline = !toop->can_be_inline_type();
3728 bool not_flat_in_array = !UseArrayFlattening || not_inline || (toop->is_inlinetypeptr() && !toop->inline_klass()->maybe_flat_in_array());
3729 if (EnableValhalla && (not_inline || not_flat_in_array)) {
3730 // Check if obj has been loaded from an array
3731 obj = obj->isa_DecodeN() ? obj->in(1) : obj;
3732 Node* array = nullptr;
3733 if (obj->isa_Load()) {
3734 Node* address = obj->in(MemNode::Address);
3735 if (address->isa_AddP()) {
3736 array = address->as_AddP()->in(AddPNode::Base);
3737 }
3738 } else if (obj->is_Phi()) {
3739 Node* region = obj->in(0);
3740 // TODO make this more robust (see JDK-8231346)
3741 if (region->req() == 3 && region->in(2) != nullptr && region->in(2)->in(0) != nullptr) {
3742 IfNode* iff = region->in(2)->in(0)->isa_If();
3743 if (iff != nullptr) {
3744 iff->is_flat_array_check(&_gvn, &array);
3745 }
3746 }
3747 }
3748 if (array != nullptr) {
3749 const TypeAryPtr* ary_t = _gvn.type(array)->isa_aryptr();
3750 if (ary_t != nullptr) {
3751 if (!ary_t->is_not_null_free() && !ary_t->is_null_free() && not_inline) {
3752 // Casting array element to a non-inline-type, mark array as not null-free.
3753 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_null_free()));
3754 replace_in_map(array, cast);
3755 array = cast;
3756 }
3757 if (!ary_t->is_not_flat() && !ary_t->is_flat() && not_flat_in_array) {
3758 // Casting array element to a non-flat-in-array type, mark array as not flat.
3759 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_flat()));
3760 replace_in_map(array, cast);
3761 array = cast;
3762 }
3763 }
3764 }
3765 }
3766
3767 if (!stopped() && !res->is_InlineType()) {
3768 res = record_profiled_receiver_for_speculation(res);
3769 if (toop->is_inlinetypeptr() && !maybe_larval) {
3770 Node* vt = InlineTypeNode::make_from_oop(this, res, toop->inline_klass());
3771 res = vt;
3772 if (safe_for_replace) {
3773 replace_in_map(obj, vt);
3774 replace_in_map(not_null_obj, vt);
3775 replace_in_map(res, vt);
3776 }
3777 }
3778 }
3779 return res;
3780 }
3781
3782 Node* GraphKit::mark_word_test(Node* obj, uintptr_t mask_val, bool eq, bool check_lock) {
3783 // Load markword
3784 Node* mark_adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3785 Node* mark = make_load(nullptr, mark_adr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3786 if (check_lock) {
3787 // Check if obj is locked
3788 Node* locked_bit = MakeConX(markWord::unlocked_value);
3789 locked_bit = _gvn.transform(new AndXNode(locked_bit, mark));
3790 Node* cmp = _gvn.transform(new CmpXNode(locked_bit, MakeConX(0)));
3791 Node* is_unlocked = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3792 IfNode* iff = new IfNode(control(), is_unlocked, PROB_MAX, COUNT_UNKNOWN);
3793 _gvn.transform(iff);
3794 Node* locked_region = new RegionNode(3);
3795 Node* mark_phi = new PhiNode(locked_region, TypeX_X);
3796
3797 // Unlocked: Use bits from mark word
3798 locked_region->init_req(1, _gvn.transform(new IfTrueNode(iff)));
3799 mark_phi->init_req(1, mark);
3800
3801 // Locked: Load prototype header from klass
3802 set_control(_gvn.transform(new IfFalseNode(iff)));
3803 // Make loads control dependent to make sure they are only executed if array is locked
3804 Node* klass_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
3805 Node* klass = _gvn.transform(LoadKlassNode::make(_gvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
3806 Node* proto_adr = basic_plus_adr(klass, in_bytes(Klass::prototype_header_offset()));
3807 Node* proto = _gvn.transform(LoadNode::make(_gvn, control(), C->immutable_memory(), proto_adr, proto_adr->bottom_type()->is_ptr(), TypeX_X, TypeX_X->basic_type(), MemNode::unordered));
3808
3809 locked_region->init_req(2, control());
3810 mark_phi->init_req(2, proto);
3811 set_control(_gvn.transform(locked_region));
3812 record_for_igvn(locked_region);
3813
3814 mark = mark_phi;
3815 }
3816
3817 // Now check if mark word bits are set
3818 Node* mask = MakeConX(mask_val);
3819 Node* masked = _gvn.transform(new AndXNode(_gvn.transform(mark), mask));
3820 record_for_igvn(masked); // Give it a chance to be optimized out by IGVN
3821 Node* cmp = _gvn.transform(new CmpXNode(masked, mask));
3822 return _gvn.transform(new BoolNode(cmp, eq ? BoolTest::eq : BoolTest::ne));
3823 }
3824
3825 Node* GraphKit::inline_type_test(Node* obj, bool is_inline) {
3826 return mark_word_test(obj, markWord::inline_type_pattern, is_inline, /* check_lock = */ false);
3827 }
3828
3829 Node* GraphKit::flat_array_test(Node* array_or_klass, bool flat) {
3830 // We can't use immutable memory here because the mark word is mutable.
3831 // PhaseIdealLoop::move_flat_array_check_out_of_loop will make sure the
3832 // check is moved out of loops (mainly to enable loop unswitching).
3833 Node* cmp = _gvn.transform(new FlatArrayCheckNode(C, memory(Compile::AliasIdxRaw), array_or_klass));
3834 record_for_igvn(cmp); // Give it a chance to be optimized out by IGVN
3835 return _gvn.transform(new BoolNode(cmp, flat ? BoolTest::eq : BoolTest::ne));
3836 }
3837
3838 Node* GraphKit::null_free_array_test(Node* array, bool null_free) {
3839 return mark_word_test(array, markWord::null_free_array_bit_in_place, null_free);
3840 }
3841
3842 Node* GraphKit::null_free_atomic_array_test(Node* array, ciInlineKlass* vk) {
3843 assert(vk->has_atomic_layout() || vk->has_non_atomic_layout(), "Can't be null-free and flat");
3844
3845 // TODO 8350865 Add a stress flag to always access atomic if layout exists?
3846 if (!vk->has_non_atomic_layout()) {
3847 return intcon(1); // Always atomic
3848 } else if (!vk->has_atomic_layout()) {
3849 return intcon(0); // Never atomic
3850 }
3851
3852 Node* array_klass = load_object_klass(array);
3853 int layout_kind_offset = in_bytes(FlatArrayKlass::layout_kind_offset());
3854 Node* layout_kind_addr = basic_plus_adr(array_klass, array_klass, layout_kind_offset);
3855 Node* layout_kind = make_load(nullptr, layout_kind_addr, TypeInt::INT, T_INT, MemNode::unordered);
3856 Node* cmp = _gvn.transform(new CmpINode(layout_kind, intcon((int)LayoutKind::ATOMIC_FLAT)));
3857 return _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3858 }
3859
3860 // Deoptimize if 'ary' is a null-free inline type array and 'val' is null
3861 Node* GraphKit::inline_array_null_guard(Node* ary, Node* val, int nargs, bool safe_for_replace) {
3862 RegionNode* region = new RegionNode(3);
3863 Node* null_ctl = top();
3864 null_check_oop(val, &null_ctl);
3865 if (null_ctl != top()) {
3866 PreserveJVMState pjvms(this);
3867 set_control(null_ctl);
3868 {
3869 // Deoptimize if null-free array
3870 BuildCutout unless(this, null_free_array_test(ary, /* null_free = */ false), PROB_MAX);
3871 inc_sp(nargs);
3872 uncommon_trap(Deoptimization::Reason_null_check,
3873 Deoptimization::Action_none);
3874 }
3875 region->init_req(1, control());
3876 }
3877 region->init_req(2, control());
3878 set_control(_gvn.transform(region));
3879 record_for_igvn(region);
3880 if (_gvn.type(val) == TypePtr::NULL_PTR) {
3881 // Since we were just successfully storing null, the array can't be null free.
3882 const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr();
3883 ary_t = ary_t->cast_to_not_null_free();
3884 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t));
3885 if (safe_for_replace) {
3886 replace_in_map(ary, cast);
3887 }
3888 ary = cast;
3889 }
3890 return ary;
3891 }
3892
3893 //------------------------------next_monitor-----------------------------------
3894 // What number should be given to the next monitor?
3895 int GraphKit::next_monitor() {
3896 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3897 int next = current + C->sync_stack_slots();
3898 // Keep the toplevel high water mark current:
3899 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3900 return current;
3901 }
3902
3903 //------------------------------insert_mem_bar---------------------------------
3904 // Memory barrier to avoid floating things around
3905 // The membar serves as a pinch point between both control and all memory slices.
3906 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3907 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3908 mb->init_req(TypeFunc::Control, control());
3909 mb->init_req(TypeFunc::Memory, reset_memory());
3910 Node* membar = _gvn.transform(mb);
3938 }
3939 Node* membar = _gvn.transform(mb);
3940 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3941 if (alias_idx == Compile::AliasIdxBot) {
3942 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3943 } else {
3944 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3945 }
3946 return membar;
3947 }
3948
3949 //------------------------------shared_lock------------------------------------
3950 // Emit locking code.
3951 FastLockNode* GraphKit::shared_lock(Node* obj) {
3952 // bci is either a monitorenter bc or InvocationEntryBci
3953 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3954 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3955
3956 if( !GenerateSynchronizationCode )
3957 return nullptr; // Not locking things?
3958
3959 if (stopped()) // Dead monitor?
3960 return nullptr;
3961
3962 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3963
3964 // Box the stack location
3965 Node* box = new BoxLockNode(next_monitor());
3966 // Check for bailout after new BoxLockNode
3967 if (failing()) { return nullptr; }
3968 box = _gvn.transform(box);
3969 Node* mem = reset_memory();
3970
3971 FastLockNode * flock = _gvn.transform(new FastLockNode(nullptr, obj, box) )->as_FastLock();
3972
3973 // Add monitor to debug info for the slow path. If we block inside the
3974 // slow path and de-opt, we need the monitor hanging around
3975 map()->push_monitor( flock );
3976
3977 const TypeFunc *tf = LockNode::lock_type();
3978 LockNode *lock = new LockNode(C, tf);
4007 }
4008 #endif
4009
4010 return flock;
4011 }
4012
4013
4014 //------------------------------shared_unlock----------------------------------
4015 // Emit unlocking code.
4016 void GraphKit::shared_unlock(Node* box, Node* obj) {
4017 // bci is either a monitorenter bc or InvocationEntryBci
4018 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
4019 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
4020
4021 if( !GenerateSynchronizationCode )
4022 return;
4023 if (stopped()) { // Dead monitor?
4024 map()->pop_monitor(); // Kill monitor from debug info
4025 return;
4026 }
4027 assert(!obj->is_InlineType(), "should not unlock on inline type");
4028
4029 // Memory barrier to avoid floating things down past the locked region
4030 insert_mem_bar(Op_MemBarReleaseLock);
4031
4032 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
4033 UnlockNode *unlock = new UnlockNode(C, tf);
4034 #ifdef ASSERT
4035 unlock->set_dbg_jvms(sync_jvms());
4036 #endif
4037 uint raw_idx = Compile::AliasIdxRaw;
4038 unlock->init_req( TypeFunc::Control, control() );
4039 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
4040 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
4041 unlock->init_req( TypeFunc::FramePtr, frameptr() );
4042 unlock->init_req( TypeFunc::ReturnAdr, top() );
4043
4044 unlock->init_req(TypeFunc::Parms + 0, obj);
4045 unlock->init_req(TypeFunc::Parms + 1, box);
4046 unlock = _gvn.transform(unlock)->as_Unlock();
4047
4048 Node* mem = reset_memory();
4049
4050 // unlock has no side-effects, sets few values
4051 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
4052
4053 // Kill monitor from debug info
4054 map()->pop_monitor( );
4055 }
4056
4057 //-------------------------------get_layout_helper-----------------------------
4058 // If the given klass is a constant or known to be an array,
4059 // fetch the constant layout helper value into constant_value
4060 // and return null. Otherwise, load the non-constant
4061 // layout helper value, and return the node which represents it.
4062 // This two-faced routine is useful because allocation sites
4063 // almost always feature constant types.
4064 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
4065 const TypeKlassPtr* klass_t = _gvn.type(klass_node)->isa_klassptr();
4066 if (!StressReflectiveCode && klass_t != nullptr) {
4067 bool xklass = klass_t->klass_is_exact();
4068 bool can_be_flat = false;
4069 const TypeAryPtr* ary_type = klass_t->as_instance_type()->isa_aryptr();
4070 if (UseArrayFlattening && !xklass && ary_type != nullptr && !ary_type->is_null_free()) {
4071 // Don't constant fold if the runtime type might be a flat array but the static type is not.
4072 const TypeOopPtr* elem = ary_type->elem()->make_oopptr();
4073 can_be_flat = ary_type->can_be_inline_array() && (!elem->is_inlinetypeptr() || elem->inline_klass()->maybe_flat_in_array());
4074 }
4075 if (!can_be_flat && (xklass || (klass_t->isa_aryklassptr() && klass_t->is_aryklassptr()->elem() != Type::BOTTOM))) {
4076 jint lhelper;
4077 if (klass_t->is_flat()) {
4078 lhelper = ary_type->flat_layout_helper();
4079 } else if (klass_t->isa_aryklassptr()) {
4080 BasicType elem = ary_type->elem()->array_element_basic_type();
4081 if (is_reference_type(elem, true)) {
4082 elem = T_OBJECT;
4083 }
4084 lhelper = Klass::array_layout_helper(elem);
4085 } else {
4086 lhelper = klass_t->is_instklassptr()->exact_klass()->layout_helper();
4087 }
4088 if (lhelper != Klass::_lh_neutral_value) {
4089 constant_value = lhelper;
4090 return (Node*) nullptr;
4091 }
4092 }
4093 }
4094 constant_value = Klass::_lh_neutral_value; // put in a known value
4095 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
4096 return make_load(nullptr, lhp, TypeInt::INT, T_INT, MemNode::unordered);
4097 }
4098
4099 // We just put in an allocate/initialize with a big raw-memory effect.
4100 // Hook selected additional alias categories on the initialization.
4101 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
4102 MergeMemNode* init_in_merge,
4103 Node* init_out_raw) {
4104 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
4105 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
4106
4107 Node* prevmem = kit.memory(alias_idx);
4108 init_in_merge->set_memory_at(alias_idx, prevmem);
4109 if (init_out_raw != nullptr) {
4110 kit.set_memory(init_out_raw, alias_idx);
4111 }
4112 }
4113
4114 //---------------------------set_output_for_allocation-------------------------
4115 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
4116 const TypeOopPtr* oop_type,
4117 bool deoptimize_on_exception) {
4118 int rawidx = Compile::AliasIdxRaw;
4119 alloc->set_req( TypeFunc::FramePtr, frameptr() );
4120 add_safepoint_edges(alloc);
4121 Node* allocx = _gvn.transform(alloc);
4122 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
4123 // create memory projection for i_o
4124 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
4125 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
4126
4127 // create a memory projection as for the normal control path
4128 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
4129 set_memory(malloc, rawidx);
4130
4131 // a normal slow-call doesn't change i_o, but an allocation does
4132 // we create a separate i_o projection for the normal control path
4133 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
4134 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
4135
4136 // put in an initialization barrier
4137 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
4138 rawoop)->as_Initialize();
4139 assert(alloc->initialization() == init, "2-way macro link must work");
4140 assert(init ->allocation() == alloc, "2-way macro link must work");
4141 {
4142 // Extract memory strands which may participate in the new object's
4143 // initialization, and source them from the new InitializeNode.
4144 // This will allow us to observe initializations when they occur,
4145 // and link them properly (as a group) to the InitializeNode.
4146 assert(init->in(InitializeNode::Memory) == malloc, "");
4147 MergeMemNode* minit_in = MergeMemNode::make(malloc);
4148 init->set_req(InitializeNode::Memory, minit_in);
4149 record_for_igvn(minit_in); // fold it up later, if possible
4150 _gvn.set_type(minit_in, Type::MEMORY);
4151 Node* minit_out = memory(rawidx);
4152 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
4153 // Add an edge in the MergeMem for the header fields so an access
4154 // to one of those has correct memory state
4155 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes())));
4156 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes())));
4157 if (oop_type->isa_aryptr()) {
4158 const TypeAryPtr* arytype = oop_type->is_aryptr();
4159 if (arytype->is_flat()) {
4160 // Initially all flat array accesses share a single slice
4161 // but that changes after parsing. Prepare the memory graph so
4162 // it can optimize flat array accesses properly once they
4163 // don't share a single slice.
4164 assert(C->flat_accesses_share_alias(), "should be set at parse time");
4165 C->set_flat_accesses_share_alias(false);
4166 ciInlineKlass* vk = arytype->elem()->inline_klass();
4167 for (int i = 0, len = vk->nof_nonstatic_fields(); i < len; i++) {
4168 ciField* field = vk->nonstatic_field_at(i);
4169 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
4170 continue; // do not bother to track really large numbers of fields
4171 int off_in_vt = field->offset_in_bytes() - vk->payload_offset();
4172 const TypePtr* adr_type = arytype->with_field_offset(off_in_vt)->add_offset(Type::OffsetBot);
4173 int fieldidx = C->get_alias_index(adr_type, true);
4174 // Pass nullptr for init_out. Having per flat array element field memory edges as uses of the Initialize node
4175 // can result in per flat array field Phis to be created which confuses the logic of
4176 // Compile::adjust_flat_array_access_aliases().
4177 hook_memory_on_init(*this, fieldidx, minit_in, nullptr);
4178 }
4179 C->set_flat_accesses_share_alias(true);
4180 hook_memory_on_init(*this, C->get_alias_index(TypeAryPtr::INLINES), minit_in, minit_out);
4181 } else {
4182 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
4183 int elemidx = C->get_alias_index(telemref);
4184 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
4185 }
4186 } else if (oop_type->isa_instptr()) {
4187 set_memory(minit_out, C->get_alias_index(oop_type)); // mark word
4188 ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass();
4189 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
4190 ciField* field = ik->nonstatic_field_at(i);
4191 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
4192 continue; // do not bother to track really large numbers of fields
4193 // Find (or create) the alias category for this field:
4194 int fieldidx = C->alias_type(field)->index();
4195 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
4196 }
4197 }
4198 }
4199
4200 // Cast raw oop to the real thing...
4201 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
4202 javaoop = _gvn.transform(javaoop);
4203 C->set_recent_alloc(control(), javaoop);
4204 assert(just_allocated_object(control()) == javaoop, "just allocated");
4205
4206 #ifdef ASSERT
4207 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
4218 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
4219 }
4220 }
4221 #endif //ASSERT
4222
4223 return javaoop;
4224 }
4225
4226 //---------------------------new_instance--------------------------------------
4227 // This routine takes a klass_node which may be constant (for a static type)
4228 // or may be non-constant (for reflective code). It will work equally well
4229 // for either, and the graph will fold nicely if the optimizer later reduces
4230 // the type to a constant.
4231 // The optional arguments are for specialized use by intrinsics:
4232 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
4233 // - If 'return_size_val', report the total object size to the caller.
4234 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4235 Node* GraphKit::new_instance(Node* klass_node,
4236 Node* extra_slow_test,
4237 Node* *return_size_val,
4238 bool deoptimize_on_exception,
4239 InlineTypeNode* inline_type_node) {
4240 // Compute size in doublewords
4241 // The size is always an integral number of doublewords, represented
4242 // as a positive bytewise size stored in the klass's layout_helper.
4243 // The layout_helper also encodes (in a low bit) the need for a slow path.
4244 jint layout_con = Klass::_lh_neutral_value;
4245 Node* layout_val = get_layout_helper(klass_node, layout_con);
4246 bool layout_is_con = (layout_val == nullptr);
4247
4248 if (extra_slow_test == nullptr) extra_slow_test = intcon(0);
4249 // Generate the initial go-slow test. It's either ALWAYS (return a
4250 // Node for 1) or NEVER (return a null) or perhaps (in the reflective
4251 // case) a computed value derived from the layout_helper.
4252 Node* initial_slow_test = nullptr;
4253 if (layout_is_con) {
4254 assert(!StressReflectiveCode, "stress mode does not use these paths");
4255 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
4256 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
4257 } else { // reflective case
4258 // This reflective path is used by Unsafe.allocateInstance.
4259 // (It may be stress-tested by specifying StressReflectiveCode.)
4260 // Basically, we want to get into the VM is there's an illegal argument.
4261 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
4262 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
4263 if (extra_slow_test != intcon(0)) {
4264 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
4265 }
4266 // (Macro-expander will further convert this to a Bool, if necessary.)
4277
4278 // Clear the low bits to extract layout_helper_size_in_bytes:
4279 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
4280 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
4281 size = _gvn.transform( new AndXNode(size, mask) );
4282 }
4283 if (return_size_val != nullptr) {
4284 (*return_size_val) = size;
4285 }
4286
4287 // This is a precise notnull oop of the klass.
4288 // (Actually, it need not be precise if this is a reflective allocation.)
4289 // It's what we cast the result to.
4290 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
4291 if (!tklass) tklass = TypeInstKlassPtr::OBJECT;
4292 const TypeOopPtr* oop_type = tklass->as_instance_type();
4293
4294 // Now generate allocation code
4295
4296 // The entire memory state is needed for slow path of the allocation
4297 // since GC and deoptimization can happen.
4298 Node *mem = reset_memory();
4299 set_all_memory(mem); // Create new memory state
4300
4301 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
4302 control(), mem, i_o(),
4303 size, klass_node,
4304 initial_slow_test, inline_type_node);
4305
4306 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
4307 }
4308
4309 //-------------------------------new_array-------------------------------------
4310 // helper for newarray and anewarray
4311 // The 'length' parameter is (obviously) the length of the array.
4312 // The optional arguments are for specialized use by intrinsics:
4313 // - If 'return_size_val', report the non-padded array size (sum of header size
4314 // and array body) to the caller.
4315 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4316 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
4317 Node* length, // number of array elements
4318 int nargs, // number of arguments to push back for uncommon trap
4319 Node* *return_size_val,
4320 bool deoptimize_on_exception,
4321 Node* init_val) {
4322 jint layout_con = Klass::_lh_neutral_value;
4323 Node* layout_val = get_layout_helper(klass_node, layout_con);
4324 bool layout_is_con = (layout_val == nullptr);
4325
4326 if (!layout_is_con && !StressReflectiveCode &&
4327 !too_many_traps(Deoptimization::Reason_class_check)) {
4328 // This is a reflective array creation site.
4329 // Optimistically assume that it is a subtype of Object[],
4330 // so that we can fold up all the address arithmetic.
4331 layout_con = Klass::array_layout_helper(T_OBJECT);
4332 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
4333 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
4334 { BuildCutout unless(this, bol_lh, PROB_MAX);
4335 inc_sp(nargs);
4336 uncommon_trap(Deoptimization::Reason_class_check,
4337 Deoptimization::Action_maybe_recompile);
4338 }
4339 layout_val = nullptr;
4340 layout_is_con = true;
4341 }
4342
4343 // Generate the initial go-slow test. Make sure we do not overflow
4344 // if length is huge (near 2Gig) or negative! We do not need
4345 // exact double-words here, just a close approximation of needed
4346 // double-words. We can't add any offset or rounding bits, lest we
4347 // take a size -1 of bytes and make it positive. Use an unsigned
4348 // compare, so negative sizes look hugely positive.
4349 int fast_size_limit = FastAllocateSizeLimit;
4350 if (layout_is_con) {
4351 assert(!StressReflectiveCode, "stress mode does not use these paths");
4352 // Increase the size limit if we have exact knowledge of array type.
4353 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
4354 fast_size_limit <<= MAX2(LogBytesPerLong - log2_esize, 0);
4355 }
4356
4357 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
4358 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
4359
4360 // --- Size Computation ---
4361 // array_size = round_to_heap(array_header + (length << elem_shift));
4362 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
4363 // and align_to(x, y) == ((x + y-1) & ~(y-1))
4364 // The rounding mask is strength-reduced, if possible.
4365 int round_mask = MinObjAlignmentInBytes - 1;
4366 Node* header_size = nullptr;
4367 // (T_BYTE has the weakest alignment and size restrictions...)
4368 if (layout_is_con) {
4369 int hsize = Klass::layout_helper_header_size(layout_con);
4370 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4371 bool is_flat_array = Klass::layout_helper_is_flatArray(layout_con);
4372 if ((round_mask & ~right_n_bits(eshift)) == 0)
4373 round_mask = 0; // strength-reduce it if it goes away completely
4374 assert(is_flat_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
4375 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
4376 assert(header_size_min <= hsize, "generic minimum is smallest");
4377 header_size = intcon(hsize);
4378 } else {
4379 Node* hss = intcon(Klass::_lh_header_size_shift);
4380 Node* hsm = intcon(Klass::_lh_header_size_mask);
4381 header_size = _gvn.transform(new URShiftINode(layout_val, hss));
4382 header_size = _gvn.transform(new AndINode(header_size, hsm));
4383 }
4384
4385 Node* elem_shift = nullptr;
4386 if (layout_is_con) {
4387 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4388 if (eshift != 0)
4389 elem_shift = intcon(eshift);
4390 } else {
4391 // There is no need to mask or shift this value.
4392 // The semantics of LShiftINode include an implicit mask to 0x1F.
4393 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
4394 elem_shift = layout_val;
4443 }
4444 Node* non_rounded_size = _gvn.transform(new AddXNode(headerx, abody));
4445
4446 if (return_size_val != nullptr) {
4447 // This is the size
4448 (*return_size_val) = non_rounded_size;
4449 }
4450
4451 Node* size = non_rounded_size;
4452 if (round_mask != 0) {
4453 Node* mask1 = MakeConX(round_mask);
4454 size = _gvn.transform(new AddXNode(size, mask1));
4455 Node* mask2 = MakeConX(~round_mask);
4456 size = _gvn.transform(new AndXNode(size, mask2));
4457 }
4458 // else if round_mask == 0, the size computation is self-rounding
4459
4460 // Now generate allocation code
4461
4462 // The entire memory state is needed for slow path of the allocation
4463 // since GC and deoptimization can happen.
4464 Node *mem = reset_memory();
4465 set_all_memory(mem); // Create new memory state
4466
4467 if (initial_slow_test->is_Bool()) {
4468 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
4469 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
4470 }
4471
4472 const TypeKlassPtr* ary_klass = _gvn.type(klass_node)->isa_klassptr();
4473 const TypeOopPtr* ary_type = ary_klass->as_instance_type();
4474
4475 Node* raw_init_value = nullptr;
4476 if (init_val != nullptr) {
4477 // TODO 8350865 Fast non-zero init not implemented yet for flat, null-free arrays
4478 if (ary_type->is_flat()) {
4479 initial_slow_test = intcon(1);
4480 }
4481
4482 if (UseCompressedOops) {
4483 // With compressed oops, the 64-bit init value is built from two 32-bit compressed oops
4484 init_val = _gvn.transform(new EncodePNode(init_val, init_val->bottom_type()->make_narrowoop()));
4485 Node* lower = _gvn.transform(new CastP2XNode(control(), init_val));
4486 Node* upper = _gvn.transform(new LShiftLNode(lower, intcon(32)));
4487 raw_init_value = _gvn.transform(new OrLNode(lower, upper));
4488 } else {
4489 raw_init_value = _gvn.transform(new CastP2XNode(control(), init_val));
4490 }
4491 }
4492
4493 Node* valid_length_test = _gvn.intcon(1);
4494 if (ary_type->isa_aryptr()) {
4495 BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type();
4496 jint max = TypeAryPtr::max_array_length(bt);
4497 Node* valid_length_cmp = _gvn.transform(new CmpUNode(length, intcon(max)));
4498 valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le));
4499 }
4500
4501 // Create the AllocateArrayNode and its result projections
4502 AllocateArrayNode* alloc
4503 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
4504 control(), mem, i_o(),
4505 size, klass_node,
4506 initial_slow_test,
4507 length, valid_length_test,
4508 init_val, raw_init_value);
4509 // Cast to correct type. Note that the klass_node may be constant or not,
4510 // and in the latter case the actual array type will be inexact also.
4511 // (This happens via a non-constant argument to inline_native_newArray.)
4512 // In any case, the value of klass_node provides the desired array type.
4513 const TypeInt* length_type = _gvn.find_int_type(length);
4514 if (ary_type->isa_aryptr() && length_type != nullptr) {
4515 // Try to get a better type than POS for the size
4516 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4517 }
4518
4519 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
4520
4521 array_ideal_length(alloc, ary_type, true);
4522 return javaoop;
4523 }
4524
4525 // The following "Ideal_foo" functions are placed here because they recognize
4526 // the graph shapes created by the functions immediately above.
4527
4528 //---------------------------Ideal_allocation----------------------------------
4636 set_all_memory(ideal.merged_memory());
4637 set_i_o(ideal.i_o());
4638 set_control(ideal.ctrl());
4639 }
4640
4641 void GraphKit::final_sync(IdealKit& ideal) {
4642 // Final sync IdealKit and graphKit.
4643 sync_kit(ideal);
4644 }
4645
4646 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4647 Node* len = load_array_length(load_String_value(str, set_ctrl));
4648 Node* coder = load_String_coder(str, set_ctrl);
4649 // Divide length by 2 if coder is UTF16
4650 return _gvn.transform(new RShiftINode(len, coder));
4651 }
4652
4653 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4654 int value_offset = java_lang_String::value_offset();
4655 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4656 false, nullptr, Type::Offset(0));
4657 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4658 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4659 TypeAry::make(TypeInt::BYTE, TypeInt::POS, false, false, true, true),
4660 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0));
4661 Node* p = basic_plus_adr(str, str, value_offset);
4662 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4663 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4664 return load;
4665 }
4666
4667 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4668 if (!CompactStrings) {
4669 return intcon(java_lang_String::CODER_UTF16);
4670 }
4671 int coder_offset = java_lang_String::coder_offset();
4672 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4673 false, nullptr, Type::Offset(0));
4674 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4675
4676 Node* p = basic_plus_adr(str, str, coder_offset);
4677 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4678 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4679 return load;
4680 }
4681
4682 void GraphKit::store_String_value(Node* str, Node* value) {
4683 int value_offset = java_lang_String::value_offset();
4684 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4685 false, nullptr, Type::Offset(0));
4686 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4687
4688 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4689 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4690 }
4691
4692 void GraphKit::store_String_coder(Node* str, Node* value) {
4693 int coder_offset = java_lang_String::coder_offset();
4694 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4695 false, nullptr, Type::Offset(0));
4696 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4697
4698 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4699 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4700 }
4701
4702 // Capture src and dst memory state with a MergeMemNode
4703 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4704 if (src_type == dst_type) {
4705 // Types are equal, we don't need a MergeMemNode
4706 return memory(src_type);
4707 }
4708 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4709 record_for_igvn(merge); // fold it up later, if possible
4710 int src_idx = C->get_alias_index(src_type);
4711 int dst_idx = C->get_alias_index(dst_type);
4712 merge->set_memory_at(src_idx, memory(src_idx));
4713 merge->set_memory_at(dst_idx, memory(dst_idx));
4714 return merge;
4715 }
4788 i_char->init_req(2, AddI(i_char, intcon(2)));
4789
4790 set_control(IfFalse(iff));
4791 set_memory(st, TypeAryPtr::BYTES);
4792 }
4793
4794 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4795 if (!field->is_constant()) {
4796 return nullptr; // Field not marked as constant.
4797 }
4798 ciInstance* holder = nullptr;
4799 if (!field->is_static()) {
4800 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4801 if (const_oop != nullptr && const_oop->is_instance()) {
4802 holder = const_oop->as_instance();
4803 }
4804 }
4805 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4806 /*is_unsigned_load=*/false);
4807 if (con_type != nullptr) {
4808 Node* con = makecon(con_type);
4809 if (field->type()->is_inlinetype()) {
4810 con = InlineTypeNode::make_from_oop(this, con, field->type()->as_inline_klass());
4811 } else if (con_type->is_inlinetypeptr()) {
4812 con = InlineTypeNode::make_from_oop(this, con, con_type->inline_klass());
4813 }
4814 return con;
4815 }
4816 return nullptr;
4817 }
4818
4819 //---------------------------load_mirror_from_klass----------------------------
4820 // Given a klass oop, load its java mirror (a java.lang.Class oop).
4821 Node* GraphKit::load_mirror_from_klass(Node* klass) {
4822 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
4823 Node* load = make_load(nullptr, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4824 // mirror = ((OopHandle)mirror)->resolve();
4825 return access_load(load, TypeInstPtr::MIRROR, T_OBJECT, IN_NATIVE);
4826 }
4827
4828 Node* GraphKit::maybe_narrow_object_type(Node* obj, ciKlass* type) {
4829 const Type* obj_type = obj->bottom_type();
4830 const TypeOopPtr* sig_type = TypeOopPtr::make_from_klass(type);
4831 if (obj_type->isa_oopptr() && sig_type->is_loaded() && !obj_type->higher_equal(sig_type)) {
4832 const Type* narrow_obj_type = obj_type->filter_speculative(sig_type); // keep speculative part
4833 Node* casted_obj = gvn().transform(new CheckCastPPNode(control(), obj, narrow_obj_type));
4834 obj = casted_obj;
4835 }
4836 if (sig_type->is_inlinetypeptr()) {
4837 obj = InlineTypeNode::make_from_oop(this, obj, sig_type->inline_klass());
4838 }
4839 return obj;
4840 }
|